Last week in his regular Exponential View newsletter Azeem Azhar described in detail how he had used ChatGPT to design a new board game that combined the characteristics of Ticket to Ride and Azul, shaping it around the idea of discovering elements, and designing selection of game characters based around historical chemists.

It's a subscriber-only post but .

Then another friend, , via GitHub Copilot, to write the code he needed to scrape a website as part of a project to build a podcast interface. As he wrote:

"Using GitHub Copilot to write code and calling out to GPT-3 programmatically to dodge days of graft actually brought tears to my eyes. I’ve coded, mostly as a hobby, my whole life – it’s a big creative outlet alongside writing – it’s so rarely felt like this. It feels like flying".

These examples brought home to me the real power of these new tools, not as generators of random boilerplate text for business letters or marketing blurb, or as complex and potentially misleading interfaces to search engines, but as collaborators in our creative activity, supporting idea generating, doing some of the low-level heavy lifting, and sitting on our shoulders like supportive angels. 

One of the things that also occurred to me reading Azeem's piece was that ChatGPT didn’t get tired.   For once, a line from the Terminator felt entirely appropriate to describe a current ML system: "It doesn't feel pity, or remorse, or fear. And it absolutely will not stop... ever, until you have finished your project!"

That’s not to say that these tools live up to the exaggerated claims being made for them, as the rather embarrassing error about the capabilites of the James Webb Space Telescope that Google made at the announcement of their Bard LLM – and the consequent $100bn drop in Alphabet’s value – demonstrated.

Of course, Bard wasn’t being malicious, or even foolish, because these tools don’t have any capacity for feeling. They can’t lie because lying is saying something false with intent, and – we can’t say this strongly enough – they have no intent. 

Pull the curtain away from GPT or Stable Diffusion and there’s no wizard, just a vast array of weightings running on a power-hungry set of GPUs. When ChatGPT engages with you it’s basically taking a drunkard’s walk through the forest of word frequencies, calling out the names of each tree as it leans on it before staggering onward. Like taking your own wine to an unlicensed restaurant with zero corkage, you bring the meaning – and because we are so good at projecting into the empty eyes of our machines (and pets.. but we can have that argument another time) we find all the profundity we’re looking for.  

Perhaps one day we will develop general AI and the machine will both know what it is saying and – crucially – know that it is a thing that is saying something to us. When that happens we’ll look back on the current fuss over LLMs the way astronomers consider astrology – there was some good data collection and analysis but the fundamental model was so disconnected from reality that it was dangerous.

But even with the current limitations, it’s clear that these tools already have a real role as well-resourced, untiring support for creativity and ideation, with the ability to smash concepts together and produce fascinating results, and that may be enough to change the way we all work, especially in the creative industries.  Just imagine what a hard-pressed producer looking for a new entertainment format could do with them.

Interventions:

When a child is typing a message, the Own It text analysis module will give live feedback on the sentiment of the message and also provides an intervention when the system determines hate, toxicity or a number of safeguarding issues. The intervention can be passive or a full-screen message – providing a nudge or friction to give the user a chance to consider their actions or whether they need help. A new addition to this system is the classification of messages related to coronavirus and isolation. The method looks for the occurrence of words or phrases concerning the issue and then measures if the emotion sentiment is anger, fear or worry. If a user were to type “I’m so lonely with this isolating”, a passive intervention would be presented supporting the child in understanding that it is okay to be worried and they’re not alone –the user is then offered a link to helpful content within the companion app. If the child was to type “I like isolation, I get to read more books”, no intervention would be presented. The nuance of when to intervene is important, it is key part of the user experience that interventions feel helpful and not an annoyance.

Collections:

Within the Own It companion app, there is a For You section that contains content which the system has determined to be important for that user. A new feature added here is Collections. These are swipe-able carousels that contain curated content concentrated on a topic. The first two collections launched with this release are The Lowdown Lockdown and Feeling Anxious or Scared?

The Lockdown Lowdown – entertaining and informative content that inspires children to live their best digital online life while in isolation – with celebrity contributions from pop stars, YouTubers and vloggers.

Feeling Anxious or Scared? – a collection of videos and articles that provide support for children during a time that they are away from their day-to-day support network of friends. Articles cover how to tackle fears and worries, de-stressing and support in understanding that it is normal to have ups and downs.

All of the new features uphold the strict data privacy requirements that have been applied to the Own It app. All of the machine learning and AI is run on the user’s device. No personal data is transferred to the ³ÉÈËÂÛ̳ and no messages are logged. All data that is generated within the system is stored and handled locally on the phone. Data privacy is regarded as sacrosanct.

Own It is available on the and . It is designed specifically for smartphones. If you have a child who has recently received their first phone or would benefit from support during these unique times, then encourage them to install Own it and let us know how you get on.

We asked presenter Radzi Chinyanganya to read to the camera the names of 12 cities, numbers from -30 to 30 and several pithy phrases to explain the temperature. The finished script sounded like this:

"Welcome to your daily weather update, let's take a look at what's been happening. In "x", residents are expecting "x", temperatures are expected to be, on average "x" so if you're heading out, remember to "x."

We used the ³ÉÈËÂÛ̳'s weather API to fill in the 'x' variable with accurate, up to date weather data from the twelve cities. You may ask at this point, why just twelve cities? To scale a demo such that a presenter can deliver a personalised weather report for any city/town/street in the world, would need advances in synthetic audio technology. When you listen to your sat nav giving you directions or you get a response to your query by a smart speaker you hear synthetic speech. Despite the explosion of investment and research using neural networks to simulate human voices, it is still challenging to replicate voices convincingly. That said, soon you won't be able to tell whether the sound of your favourite celebrity is synthetic or authentic. For our experiment, we decided to use Radzi's real voice, instead of a sub-optimal digital version that would've broken the illusion of the experience.

Take a look at and see the results for yourself. Select your favourite city and get a personalised synthetic video report based on real-time weather data.  Please note this demo only works in Google Chrome and other Chromium based browsers such as Brave, Opera and the new Microsoft Edge.

Safeguarding Trust

Conducting experiments with such contentious technology for a responsible public service broadcaster is tricky. Thorny issues of trust and editorial norms quickly come to the surface.

Trust with audiences is foundational to the ³ÉÈËÂÛ̳. It is clear that viewers watching or listening to fake content that, on the surface, appears authentic risks reputational damage. However, that's not to say there are no circumstances where the use of synthetic media could improve the audience offer without sacrificing trust. A lot depends on us being honest and clear with the audience about what they are getting, an editorial principle that the ³ÉÈËÂÛ̳ is used to applying in all sorts of contexts. The use of synthetic media in a news context has, as outlined above, the potential to be de-stabilising, especially in an era of 'fake news'. However, in a different context, like our weather report demo, it is unclear that audiences would be troubled if digital avatars were delivering a weather report. Given the growth of digital assistants and the industry drive for greater personalisation, perhaps there will be an expectation that a video response to a query will be digitally generated.

Another factor to consider that may help with trust would be audience markers. Similar to many online chatbots that use robot emojis to convey to the audience that they are speaking to a machine, not a human, it is entirely possible to use similar visual markers to communicate to viewers that a piece of content is computer generated. In this context, with the added safeguards in place, the growth of synthetic visual media seems plausible even for a responsible public service broadcaster.

The second and perhaps more intriguing issue that arises when thinking about synthetically generated media is editorial. Take the weather demo, even the most generous critic would concede that it's a bland weather report. The storytelling flair and creativity presenters bring to enrich a piece of content is completely lost in this dispassionate demo. One of the significant challenges in a world of computer-generated media will be working out how to create dynamic, imaginative content in a personalised way. Or perhaps to work out how to use technology to deliver the bits of the presentation that are bland but labour intensive and thereby give our talented storytellers more time and space to create valued content in tandem. That's not to say that bland content is an inevitability - the emerging field of AI personality designer could perhaps lead to hugely creative synthetic experiences.

So, back to our original question, can synthetic media drive new content experiences? Yes – I believe it can. Currently, the costs to deliver high-grade synthetic video are prohibitively high for ordinary consumers. As the tools become increasingly commoditised, consumers creating quality synthetic experiences at a low price could conceivably unleash a new model of storytelling. You soon imagine a future where a photorealistic human-like digital character can be made to do anything from reading out the football results to delivering a physics lesson.

At a time when the world is increasingly troubled by authentic false content, the challenge will be, in short order, to work out how to prepare for this storytelling paradigm shift.

Four hundred guests, a score of top-notch speakers, two rooms of cutting edge technology demonstrations, two programme recordings, and an orchestra playing music composed by a machine. All in a day’s work for this week’s in the Radio Theatre at Broadcasting House, organised by the ³ÉÈËÂÛ̳ Blue Room.

In this, the third year of our AI themed event, the focus was on responsibility, and how we can tackle the many issues raised by the increased use of AI and ML. Emphasising the ³ÉÈËÂÛ̳’s commitment to lead by doing, Chief Operating Officer Grace Boswood began the event by announcing:

• The ³ÉÈËÂÛ̳’s first ethics guidelines for in-house machine learning

• The publication of the outcomes of significant research examining global attitudes to technology and society

• New data apprenticeships the ³ÉÈËÂÛ̳ is collaborating on with a range of other organisations.

In the first panel, host Tina Daheley discussed the twin challenges of responsible data handling and ensuring machine learning is fair to those affected. With panellists Sana Khareghani from the Office for AI, Sandra Wachter from the Oxford Internet Institute and Indra Joshi from NHSX, the discussions ranged from minimising bias to enlisting public support through demonstrating tangible benefits.

The next session – a live recording of Radio 4’s ‘The Media Show’ - focused on machines making decisions for our children, and considered whether it was enough to educate young people to be ‘algorithm-literate’ or if more regulation was needed.

Spencer Kelly, from the technology programme Click, was brave enough to provide on-stage demonstrations of the ³ÉÈËÂÛ̳’s new , which helps children negotiate the online world from the time they get their first smart phone. The second demonstration was a synthetic video weather forecast. The weather forecast can generate hundreds of flawlessly realistic versions personalised by locality or other variables. It prompted a serious debate about whether the ³ÉÈËÂÛ̳ should ‘fake’ its reports in this way. The view in the theatre and in the ‘Digital Human’ themed Fair outside seemed to be as long as it’s made clear that the content is synthetic then it could be a powerful tool in our push to provide more personalised services.

At the Fair, automated journalism rubbed shoulders with avatars, we made eye contact with mini-robot holograms, competed with machines to spot news stories and watched an ‘ML’ cocktail-dispensing robot scan our faces and decide which drink we ‘deserved’. The conference hashtag - #bbctechconf - was used enthusiastically throughout the day with one delegate observing, “my face broke the #gin machine. It looks at you and predicts your poison. It poured me a triple sloe gin before the human leapt in to reset saying something about how it’s not supposed to do that...!

The ³ÉÈËÂÛ̳’s New Experiences UX team were on hand with some clever future-facing journalism on a news stand that threw us forward to 2040 with the help of an unashamedly retro flat-capped paper-seller, ensuring we didn’t leave without news that the ³ÉÈËÂÛ̳ had created the digital media currency the ReithPiece, while the self-care industry had dangerously depleted the world’s Sandalwood forests.

One of the most important things the conference did was to emphasise the notion that machine learning is about more than just a dense concoction of data and maths but also involves creativity, as the ³ÉÈËÂÛ̳ Philharmonic Orchestra performed a three-part work jointly composed by Robert Laidlaw and an AI ‘composer’, and forced us to ask ‘who gets the credit here?’

After the performance we entered a disinformation dystopia in which ³ÉÈËÂÛ̳ News’ Editorial Director Kamal Ahmed dished out questions to a range of guests. Disinformation is an area the ³ÉÈËÂÛ̳ is spending a lot of time on – not least through its ‘Beyond Fake News’ work, and it was valuable to hear from World Service Group Director Jamie Angus. The panel proffered suggestions on tackling disinformation such as adding ‘friction’ to the sharing of suspect content and getting into schools early with effective education about the dangers of disinformation; but the danger remains ever present and it was clear that none of the panellists believe our current defences are strong enough. Are broadcasters operating a Tetris model in a Minecraft world? mused writer Rachel Botsman.

Thanks to Cassian Harrison from ³ÉÈËÂÛ̳ Four we were also treated to a sneak preview of Ian Hislop’s programme about the history of ‘fake news’ due to air on the channel that night (spoiler: it was brilliant). And while we were in the mood for some comedy, perfectly punctured any danger of a mid-afternoon conference slump with a tech themed set. Who amongst us hasn’t chuckled at a social media pal with a chronic case of ‘Boomerangitis’?

We closed the conference with a session recorded for the ‘Beyond Today’ podcast, and Jamie Bartlett, Natalie Cargill and Stephanie Hare offered thought-provoking views on everything from how facial recognition should be regulated to how authorities might be more generally held to account about the more invasive uses of AI.

Regulation had been a recurring theme of the conference, with Facebook’s Simon Cross and Children’s Commissioner for England Anne Longfield amongst those suggesting we needed more rules to follow. There was also a clarion call for all of us to speak out when we see these powerful technologies used in ways we’re uncomfortable with or that we feel might discriminate or damage, because if we don’t, it’ll be too late. We need to educate ourselves and those using AI need to work strenuously to communicate their decision-making. As Jamie Bartlett commented, AI is starting to put ethical distance between those of us building it and those of us affected by it.

In her opening speech, Grace Boswood had confided her childhood dream of being an air hostess. We were honoured that our 400 guests decided to fly with us on a damp Monday morning, touching down eight hours later having listened, watched, learned, questioned and networked. And wanting to fly is evidently something women in technology hold in common. Ada Lovelace, who we were preparing to celebrate the following day, studied bird anatomy and mechanics as she tried to build a machine to shake off the surly bonds of earth. AI will offer us new ways to get airborne – it might even help us finally get our hands on those Jetson flying cars. To stretch the metaphor to breaking point, the conference helped us think about how we police those skies, how we make flight available to all and how we avoid flying too close to the sun. As Nejra van Zalk from Imperial College observed “All of us are part of a massive social experiment and we don't know what the outcome will be.” We have to work together to make sure we fly safely.

Own It consists of a custom keyboard and companion application. The keyboard features technology and machine learning that can assess, intervene and provide help for areas including bullying, hate, toxicity, data privacy and emotions. Once installed, the Own It keyboard becomes the default for all text input fields, covering apps and web applications. This allows the system to provide in-the-moment support, in many cases, before a message is sent – encouraging children to stop and think before they share.

The Own It companion app allows the user to see how they are using their device alongside being able to keep a diary of their feelings. Content is presented that aids children in understanding the available insights and guides them to develop resilient behaviours.

The Own It app was a complex product to develop with many technical challenges to overcome. What follows is a whistle-stop tour through six of the key ones:

1: Providing live feedback within a keyboard

A major feature of Own It is that the keyboard can provide the right support at the right time. To enable this we needed the text to be analysed as it is being typed. A text analysis module, powered by machine learning models, assesses the text for hate, toxicity, emotions, privacy data and a series of safeguarding issues. The system will instantly respond with an intervention where necessary and will also keep the user informed on the emotion sentiment of the message.

When children are in human social situations, they may step over an adult imposed red-line or social norm – but they will, usually, quickly realise that there is a problem because they can see reactions on the faces of the adults or their peers. Once the issue is identified, the child can adjust their behaviour to quickly fit into the norm for the situation. When engaging in online activities, there is little in the way of visual cues, red lines can be crossed in terms of behaviour and the user may not realise until much later. The Own It keyboard is seeking to fix this via the means of a feedback mechanism. By putting a face on the keyboard, the user can see how the message may be received. The face will display how the recipient may interpret the message – this nudges the sender to think about the tone of the message and consider adjusting it to something more palatable.

The face takes a feed from an emotion sentiment classifier that has been developed for this task. The classifications have been derived from the Warwick-Edinburgh Mental Wellbeing Scale, a measurement method that has underpinned much of the project design.

When issues with hate or toxicity, privacy or safeguarding are determined, the keyboard will present a related intervention. This is displayed in one of two forms.

A passive intervention will give the child information but without interrupting the flow of input – for example, “Are you sure you should be sharing your phone number?” – We don’t know who the recipient of the message is going to be, so raising the question to the child, allows them to think and judge if it is the right course of action.

A full intervention covers the keyboard entirely. The user is forced to acknowledge the issue by either dismissing the intervention or finding out further information. Own It will never stop a message from being sent, it just seeks to ensure the child is supported as fully as possible.

2: On-device encapsulation

During the discovery phase of the project, we built a proof-of-concept text analyser. This allowed us to begin to experience, measure and comprehend how real-time analysis could work. For the first prototype, a cloud-based solution was implemented - 3.5 GB of machine learning models and word-embeddings hosted on a server. For both data security and speed of response reasons, we wanted to move all text analysis to function entirely on the users’ device.

It seemed like a stretch, but it was essential to explore whether we could achieve the objective of miniaturising the model ensemble to the extent required. An iteration of our proof of concept applied FastText compression to the text classification, producing a first on-device build that totalled less than 40mb, only 1% of the original size. Some accuracy was lost with the models, but we had proved that a serviceable model ensemble could be placed onto a smartphone to achieve the product aspirations.

When moving to production, different compression methods were applied to the models that allowed the size to be reduced to less than 20mb. Further work on the datasets and model design meant that the final f1, recall and precision scores were better than the results from our original 3.5 GB model set.

A key element that enabled the process was to establish a workflow that enabled the model ensemble to be constructed and then ported to both Android and iOS. On-device machine learning is still in its infancy, certainly in terms of natural language processing – however, a route was found using a Keras CNN base model. The Keras model can be transformed to function with CoreML on iOS and with Tensorflow Lite on Android. Developing and testing on the single Keras model reduced the amount of resource required to ensure the efficacy of the system, as well as providing parity between the platform experiences.

3: Ensuring the quality of feedback

The data sets used to train the machine learning models - for hate, toxicity and emotion sentiment - consisted of full sentences. The sort a person would send in a message or put into a note. However, when a user is on a device, inputting text, the sentences don't just appear whole - they are typed in sequentially. This is problematic for a system that is making a prediction on an input, when that input is incrementally changing.

The initial builds and prototypes of Own It provided an experience that showed a confused emotion sentiment feedback mechanism, certainly when the early part of a sentence was being typed. We needed a method to neutralise those first few entered words of a sentence while they were being entered, only engaging the text analysis when there was enough context to make a reasonable prediction on the emotion sentiment. Word or character count was a far too simplistic method, giving poor results. Analysis of the grammatical construct of sentences was a step too far - certainly for a project that already had many complex moving parts.

The implemented solution to this problem was surprisingly simple but provided a significant improvement to the user perception of the issue. 1/2 million messages, mostly from young people were analysed and a collection was made of the most common 'sentence starts'. That is, words that are present in a sentence before the object, subject or other context is imparted. For example:

• What was the...
• Can I have a…
• Thank you for your…

The list was filtered down to contain only the most common ‘starts’, eventually totalling less than 2000. When a user types a message that matches or partially matches anything in the list, the text isn't analysed and therefore the visual feedback is neutral. Because the list only contains partial sentences, the text analysis and related responses kick in as soon as the emerging sentence begins to have meaning. This function reduced the occurrence of early and uninformed analysis by an amount that has all but taken the problem off the table. While restricting variances in sentiment response early in most sentences, the system still allows short or uncommon sentences to be analysed and responded to. The user experience is much more consistent, allowing trust to grow in the asserted feedback on display.

4: Restricting the occurrence of bias within results

All data contains bias – much of the training data for Own It came from human sources, so the biases are innate. The challenge for Own It was to counter the bias, both in the training sets and by using pre-processing methods, so that the outcomes are as unaffected as possible.

One area of potential bias came up with the emotion sentiment classification, where the displayed representation is a facial reaction to the emotion of the message. There was a worry that protected characteristics, such as race, gender and religion could have a possibility of affecting the results and provide responses that could cause offence (a misplaced angry or frightened face). As the key function of the emotion sentiment classifier is to determine only the emotion of a given sentence, the presence of any protected characteristic should not affect that response. With this taken as a rule, we needed to remove the protected words before analysis – without affecting the construct of the sentence. A ‘Bias Neutraliser’ list was drawn up containing the characteristics and grammatical variations.

This Bias Neutraliser list is parsed when the message is being pre-processed ahead of submission to the text analysis models. If a word or phrase matches the list, then it is neutralised within the sentence. To retain the sentence structure, the matched word is replaced with a neutral word – in the implementation, literally the word “neutral”. So, “the Christian walked down the stairs” is passed through the analysis as “the neutral walked down the stairs”. In this case, the word ‘Christian’ is given no opportunity to influence the emotion sentiment results.

The implemented Bias Neutraliser pre-processing affects only the system responses and provides a layer of support for the emotion sentiment analysis, allowing it to focus on its core purpose.

5: Developing Children's focussed keyboard

When trying to make a keyboard that is to be used by children – it isn’t just all of the basic elements that need to be in place, they each need to be developed with a young audience in mind. Autocorrect, autocomplete and next-word prediction must perform to a high standard AND be relevant for the users.

To support these core keyboard functions, we developed a children’s dictionary. An analysis of messages written by young people produced a word density map of the most used words and their frequency of use. The result was compared to a standard adult-focussed list and merged to create a core system dictionary that ensured the broadest range of users could benefit.

With improvements driven by words commonly used by young people, there was still an omission related to the uptake of new online language by young audiences who engage with the plethora of available platforms. New words and phrases emerge and are used within messaging on specific platforms, a growing expansion of the lexicon driven by new technology and its users. Whether it is yeet, thicc, poggers or oomfs (look them up - the etymology of the chat speak area is a fascinating story, one not for this blog, however.). The Own It keyboard is designed to work on all media platforms seeking to make the experiences as friction-free as possible. For this to work well, the functions need to be cognizant of these new words. A review of this ‘chat-speak’ allowed us to create a glossary that informed an expansion of the dictionary, giving it a wide enough set if terms for now, but this will be an ongoing journey.

6: Ensuring data privacy

Data privacy within the Own It experience has been a primary objective. It is the reason we have encapsulated the machine learning models onto the device and is also the reason we don't key log any of the data. When a message is typed, it is passed for analysis. Once a user sends or deletes their message - no record of it is stored. A data point is generated for each message sent. This contains just the sentiment of the message and a flag for any interventions types that may have occurred. This data point is placed in a secure data store on the device and can only be accessed by the companion app - the system will analyse these data points, looking for insights which can be presented to the user via support content.

The setup of the Own It system gives the machine learning a closed loop. A user can report false positives. If they receive an intervention, they can report that it was incorrect - by investigating the patterns of these false positives, we can work out where some of the classifiers aren't working as well as expected and put effort into finding and shaping further datasets that can help to improve the accuracy of these models.

Outside of this user feedback loop, the Own It team will be looking to improve the datasets and implement new techniques to increase the accuracy and performance of the full machine learning ensemble.

The advancement and proliferation of AI has created a window of problem-solving capabilities. With these new capabilities, comes a need to act both ethically and with responsibility. Within the Own It product, we have endeavoured to employ machine learning for good and treat personal data as sacrosanct. It is crucial for users to have confidence that the system is there to support them and is worthy of trust. Within the companion app, we have content that describes how Own It works. Written in an easily digestible form, it seeks to add a layer of transparency that will grow trust in Own It and that it is there to provide dedicated help to the individual.

 

To develop and build the Own It product machine learning and analysis elements, the ³ÉÈËÂÛ̳ worked in partnership with Switzerland-based Privately. Some of Privately's previously developed models were enhanced and encapsulated onto devices, while new models and related technologies were co-developed and implemented. The audience facing elements of Own It were developed and built in partnership with Glasgow’s Chunk Digital.

Own It is now available on the and . It is designed specifically for smartphones. If you have a child who is about to receive their first phone or has recently done so, then get them to install Own It and let us know how you get on.

Challenge	Machine Learning solution	Editorial solution	When we collaborate
How do we ensure curation is a good experience for users?	We consider many different measures of success: accuracy, diversity, recency, impartiality, editorial priority.	Traditionally on an editorial team, a journalist would research a story, discuss how it might be covered and compose the story itself to make it compelling.	The data scientists get a rich understanding from editorial of the different trade-offs between these measures of success. Deep domain knowledge.
How does recency impact curation of content?	We include publication date as a feature in our models. We sometimes try and optimise for recency, showing people more current content in some situations.	One of the challenges is that once that work is done it is fairly hard to bring the editorial creation back to life, especially for evergreen content. This is one of many examples that ML recommendations could help with, by surfacing this content in the most relevant time according to the user’s experience or history.	By working together we’re able to identify how to make decisions about which pieces of content are evergreen and suitable for recommendation, and which pieces have a limited shelf-life and shouldn’t be presented to users beyond a certain point.
How does the ³ÉÈËÂÛ̳ ensure impartiality?	We use  to understand if our model is giving unbiased results.   Good practice in machine learning make sure that we’re using unbiased training data.	Editors, journalists and content creators make a concerted effort to ensure that a range of views and perspectives are shown within a piece of content or across several pieces of content(within a series for example)	We combine our good practices with domain knowledge from editorial. We use techniques like human-in-the-loop machine learning, or semi-supervised learning to make editorial’s lives easier, and apply their knowledge at massive scale.   ML helps editorial identifying those pieces of content that show a breadth of views.
How we ensure variety within content serving?	We construct mathematical measures for  We include these in our machine learning optimisations.	Editorial staff responsible for curation ensure a breadth and depth of content on indexes, within collections etc	We learn about the differences between our different pieces of content. Working together we’re able to determine if our recommendations offer an interesting, relevant, and useful journey for the user.    The ³ÉÈËÂÛ̳’s audio networks feature different output and tone of voice. ie. Radio 4 has a very different ‘flavour’ to 6Music. Consequently network can be used to ensure variety in results.
How do we avoid legal issues?	We are given a checklist, and we check the items off. We get told that there are things “we can’t do for opaque legal reasons” but never really understand why, and limit the functionality of our solution.	Editors, journalists and content creators have to attend a mandatory course relating to media law, so that they have full knowledge about issues such as contempt of court, defamation and privacy. An editor will sign off content to ensure that content is compliant with legal requirements.	By talking to legal advisers we can build business rules to minimise the risk of legal infractions.    Close collaboration with editorial means we gain a deep understanding of the potential problems ahead at an early stage. We build with awareness of these concerns, and with that awareness build a solution that is high quality from both a technical and editorial point of view.
How we handle editorial quality?	We build and refine a model using data science good practices, and then turn it over to our editorial colleagues. They then decide if the results are good or not.	When editors curate they can choose content that is relevant, interesting and of good quality.      Recommendations present a specific editorial challenge, in that recommenders can surface content that is not the best of our output.	In ³ÉÈËÂÛ̳+ we prioritised content that we knew would suit the environment in which it appeared: standalone, short-form videos, appearing in a feed, from digital first areas such as Radio 1, The Social, ³ÉÈËÂÛ̳ Ideas etc   Including editorial throughout the process means that they teach us about what is important in the results, so that data science understand the real problems that we’re trying to solve.   We fail quickly, and learn quickly, getting to a better quality result.
How we learn from our audiences? Accuracy/user generated content?	Measure user activity with the products, and construct measurements of engagement.   Building implicit and explicit feedback loops. An explicit feedback loop is having a“like” button, an implicit feedback loop is determining a way to measure when something has gone wrong, like bounce rate or user churn.	We monitor feedback and analyse stats to build a picture about how our audiences engage with our content.	We work with editorial to understand the insights we get from data. They help rationalise the behaviours that we see in the data. They also teach us things that we should look for in the data.
How we test recommendations	A mixture of offline evaluation metrics (e.g.testing against a known test set of data), and online evaluation metrics (e.g.A/B testing)	Traditionally: We monitor feedback and analyse stats to build a picture about how our audiences engage with our content.	The editorial lead works with data scientists on the composition of the recommender. The results are then reviewed by the editorial lead and to obtain a variety of opinions the results are reviewed by more editorial colleagues.    More on quantitative testing here .   The rich editorial feedback lets us understand where our model could be better and make improvements.

We’re big believers in cross-disciplinary collaboration. As we’ve touched on in this article the ³ÉÈËÂÛ̳ has a lot of uniquely complex problems to solve in this space. This collaboration is essential if we’re going to continue to deliver value to the ³ÉÈËÂÛ̳’s audience using data.

If you are curious about this collaboration and would like to know more in depth about how we work, leave us a message and we will be happy to get back to you.

Also, we are hiring .

 “All around the world, fake news is now the poison in the bloodstream of our societies – destabilising democracy and undermining trust in institutions and the rule of law” - Speech by Tony Hall, Director-General of the ³ÉÈËÂÛ̳ - Lord Speaker Lecture - Wednesday 20th March 2019.

Propaganda, deception, suppression of free speech, have all been enduring issues for every society, but in recent years terms like ‘fake news’ and disinformation have been heard in public discourse with alarming regularity. So, what is happening to make it a live issue for news organisations? Can anything be done to push back against the wave of disinformation? What type of interventions are needed? Can ML help tackle disinformation?”

The latest fireside chat was hosted by ³ÉÈËÂÛ̳ Technologist Ahmed Razek. The panel line-up for the evening featured Sam Jeffers (Who Targets Me?), Dr. David Corney (Full Fact), Magda Piatkowska (Head of Data Solutions, ³ÉÈËÂÛ̳ News), and Jon Lloyd (Mozilla Foundation).

Ahmed Razek kicked off by setting out First Draft News’s .

The world of misinformation is complicated. Do people actually care about having real news that challenges them?

Sam Jeffers feared that we only think about disinformation a bit, not enough. Who Targets Me is trying to normalise people’s understanding. We see strange things from time to time that deserve explanation. There is a growing community of people being confronted with misinformation. There is a need to help people find trust signals to help them differentiate between trustworthy and untrustworthy content. If we can be more transparent, we can make more of the trustworthy content more trusted. Madga Piatkowska stressed the need for developing data solutions without hurting people. The intent behind publication and content is an important aspect. Satire is not true and “facts” are not always facts - not everything is intended to misinform.

Jon Lloyd, referring to his advocacy work at Mozilla, thought it is all too easy to fall into the trap of talking about fake news. Disinformation is affecting every aspect of our daily lives now. This is a sociological problem, spanning human rights, political and health arenas, and so on. Companies behind tech need to be looked at closely. The public is coming along with Mozilla on disinformation as a term. In the US, a recent survey showed that people are more concerned about disinformation than terrorism.

We are discussing how ML can tackle disinformation. Jon has advocated for one simple tech change - The Guardian’s data labelling feature.

Jon shared his relevant experience of proactive media action in the face of disinformation. The Guardian noticed a lot of traffic on a 2013 article quite suddenly (it was an old article). Traffic was coming from a Facebook group which was posting a lot of Islamophobic content. The Guardian knew that people were not paying attention to the date of the article so they tweaked the metadata to make the date immediately noticeable to the reader. Taking a human-centric approach to do what was in their power - to change what was happening. Lots of blame is reflected on the media for not doing enough. There are more sophisticated threats now, more authentic accounts spreading misinformation. We need more transparency on organic content (user-generated content). It is necessary to work with researchers to set a baseline of what excellent looks like and to assess against that baseline. Jon encouraged Technologists to support transparency efforts to get to excellent.

The nature of elections is changing. What do technologists and journalists need to prepare for going forward?

Sam thinks that we regulate tightly in the UK. Who Targets Me is interested in people being able to prove who they are, particularly if they are running large amounts of political advertising. Some special cases deserve anonymity but an individual, group or organisation should generally be able to stand behind what they put out. Do people really understand why they see a particular message or content, based on the data collected on them? Democracy is about debate and collective decision - we need to explain modern campaigning approaches and raise faith in how elections are run. Facebook doesn’t expose information about targeting - what data is used to reach particular people. Social media tools allow for the circumvention of conventional electoral practice.

Can the panel share some insight into the fact-checking process?

Magda shared observations of ³ÉÈËÂÛ̳ News’ work with Reality Check journalists. ³ÉÈËÂÛ̳ News has a role in transparency, in explaining to the audience what happens. Most people don't understand what targeting actually is. It is very important that we do explain. Sam maintains that Facebook is an interesting dilemma as they have done more than other platforms, but take multiple-times more money for this type of advertising. Google and YouTube transparency tools are polluted; they are not clear on how often they are updated and they are messy.  

David Corney shared useful insights into Full Fact’s fact-checking carried out by journalists - checking claims by influential people that may be misleading or easily misinterpreted by the audience. The fact-checking journalists publish a fact check; a piece summarising the full story after doing the research that the audience does not have time to do. A smaller communications team checks when these claims are being re-shared.

Newspapers are asked to publish corrections but regularly decline the invitation. Full Fact’s automated fact-checking team is a team of technologists working to support the fact-checkers and the organisation’s communications team, using ML. Prototype software to do fact checks of full sentences is being developed and refined. Algorithms will find a series of data and check if a claim is true or false for more straightforward claims. Full Fact recently received Google funding to build a better claim detection system. Concrete claims will be stored, labelled and tagged. This will allow a wider range of media and free up fact-checkers.  The potential dangers of disinformation are making the ³ÉÈËÂÛ̳ risk-averse, and in journalism, this is a problem as the speed of publication is important. Increasingly, it is not that we have a problem with the process, but that we have a problem with the competition.

Recent ³ÉÈËÂÛ̳ News research suggested that, “In India, people are reluctant to share messages which they think might incite violence, but feel duty-bound to share nationalistic messages”. What does the global view of the impact of disinformation look like? What is the non-western perspective?

Magda said that different patterns and preferences are witnessed across the globe. Long journalistic tradition is not the case everywhere. Literacy challenges, accessibility to online content, and ability to scan and consume it are prevalent in certain regions. We must also consider the impact of government and propaganda in certain areas. Jon added that we could end up creating policies that are difficult to enforce on a global scale. Magda thought that we needed to pay attention to where the tech is going to grow, e.g. China, as data will impact the way in which disinformation will spread in those regions. Are scoring systems a viable tool to rate content? David felt that 20-30 years ago when content was primarily garnered from newspapers and TV news, editorial teams acted as gatekeepers. That role has been somewhat demolished by social media and citizen journalists who spread their stories. We need something to point us to the stories that are worth paying attention to. If the algorithm gets it wrong, automation will be damaging.

What role do algorithms have to play?

Ahmed moved the conversation along to the subject of recommendation algorithms. Sam pondered, “When it strikes you that you see a Facebook ad and you click through and then you are recommended other pages, how quickly can that send you in more radical directions than you were expecting - to some strong content?”.  Regarding recommendation engines built a while ago, we don’t really know where the accountability lies. Do we understand people’s information diets?  People are consuming lots of stuff from a particular perspective and wonder how they got there. Magda argued that if you really rely on ML you have to take into account that your algorithm learns from people’s behaviour. That behaviour is not always good for them; they sometimes have poor information diets. This is when we analyse what it means to be informed by editorial and policy strategy as well as tech.Start simple so recommenders are not too complicated, and so we can assess if we are hurting the audience. If you put more interesting content in front of people, they do engage. Take the audience and journalists on a journey.

David thought that algorithms have a tendency to go towards most extreme content. Algorithms do give us relevant recommendations but sometimes get it wrong. Recommendations systems can look to the authority of sources rather than recency. Jon reflected that ultimately we need transparency. Platforms say they are making tweaks and fixes that cannot be proven. We are supposed to take at face value these companies who have profit and expansion at their core. Sam agrees - if a business model is totally dependent on the algorithm, and platforms are optimising for engagement and the scale is huge, switching them off is a massive decision. Magda reminded us that this should be the responsibility on the supply side of the content also. 

In the UK, a recent report on misinformation by the Commons Select Committee suggested that a new category of tech company should be formulated, “which tightens tech companies’ liabilities, and which is not necessarily either a ‘platform’ or a ‘publisher’”.

There has to be a system, according to Magda. There is no one object of regulation - not platforms, or media, or government. It is the responsibility of the system. All parties have their part to play. Media has a role to educate. Sam restated Who Target Me’s interest in radical transparency around political advertising. There may need to be a product suite for solving all the different problems. Different tools are required to do different jobs, and a market created in tools to help you to understand.

A team of researchers at the Allen Institute of AI recently developed Grover, a neural network capable of generating fake news articles in the style of human journalists. They argue they are fighting fire with fire because the better Grover gets at generating fakes, the better it will get at detecting them.

The ability to generate text did not worry David, the problem is getting the content into a platform where people start believing it. The story is not a problem in itself. Madga argued that it depends on intent. Who is behind it, using for good or for bad?

There has been some hype around both deep fakes and shallow fakes. A recent example of the shallow fake was the slowed-down video of Nancy Pelosi which made her appear to be disoriented. This video was subsequently retweeted by the President of the United States. There was no ML required here, this was basic video manipulation that has a profound effect.

Jon believed that a picture is worth a thousand words. Video even more so. Preparedness is better than panic. We should be more concerned about recommendation algorithms, methods of verification and systems to flag false content. To unfollow YouTube video is a long process. Changing policies is one thing. Responsible behaviour on the part of companies is not a zero-sum game. Sam thought that political video is shallow-fakey anyway. It’s telling a story via the use of selective information. David advocated that it is worth considering radical options like massive regulation. Magda thought trust will become such a big thing; the brand association with factual content. And she foresaw a decline of the not-so trusted-brands. Jon reflected upon transparency transformation in the food and fashion industry but also recognised that there is no silver bullet. It will require a coordinated effort offline as well as online, and not just tech. While financial incentive remains for companies, it won’t happen on its own. Sam added that we can use this tech to make good democratic strides forward also.

Huge thanks to Ahmed Razek and the panel for delivering another engaging fireside chat on a very hot topic. The conversation around fake news, misinformation and disinformation is multi-faceted. As the ³ÉÈËÂÛ̳, we need to keep reminding ourselves and others that the problem is not just about journalism. The impact of misinformation reaches far and wide and needs to be considered from societal, policy, tech, humanitarian and public trust perspectives. And so we, along with other organisations, are taking a deeper look at what is happening in these areas. There is lots of great ongoing work in ³ÉÈËÂÛ̳ R&D, ³ÉÈËÂÛ̳ News, and elsewhere in the organisation. The ³ÉÈËÂÛ̳ provided feedback into the Disinformation and “Fake News”: Final Report (February 2019). Director of the ³ÉÈËÂÛ̳ World Service Group, Jamie Angus, subsequently confirmed that the World Service would take the lead in addressing the ‘Fake News’ threat making use of its 42 language services, knowledge on the ground and ³ÉÈËÂÛ̳ Monitoring to spot harmful examples and expose emerging patterns. To echo Magda, we must progress in a way that is not harmful to our audience.

The problems created by online misinformation are not trivial, and the threats to society are genuine. Take the recent emergence of the anti-vaxxers movement where false information about the dangers of life-saving vaccines targeted at a newly receptive and sizeable audience across social network platforms led some parents to put their children at medical risk. Though the dissemination of this material is not illegal, it is undoubtedly harmful.

Big Business

Influencing or subverting democratic norms isn’t just about being able to manipulate people; it’s also big business. There is a lot of profit to be made by telling people what to think, and social media has become the cheapest way to accomplish this.

Social media and video hosting services are playing a significant role in circulating misinformation, both on public channels like Twitter and over encrypted messaging services like WhatsApp. There have been worldwide calls for media and technology companies to take more responsibility for content hosted on their platform.

In the UK, a recent report on misinformation by the Commons Select Committee suggested that a new category of tech company is formulated, “which tightens tech companies’ liabilities, and which is not necessarily either a ‘platform’ or a ‘publisher’”. At the same time, the UK Government plans to consult on its ‘Online Harms’ White Paper, a joint proposal from the DCMS and the Home Office. A public consultation on the plans is currently running for 12 weeks until July 1st 2019.

Regulation

Germany recently implemented the Network Enforcement Act, which forces technology companies to remove hate speech, fake news and illegal material or risk a heavy fine. Notwithstanding freedom of speech concerns, it is not clear that the law is working as intended, despite placing a heavy burden on the platforms.

Lawmakers are clearly dissatisfied with the status quo, but it remains unclear what new types of responsibility will be placed on online services as a result. Conjuring up workable law to control what appears online is hard, and any regulation is unlikely to be universally acceptable.

Outside of regulation, there is growing consensus around the need for greater media literacy campaigns. It is vital that we teach people of all ages to be critical consumers and sharers of information, especially in the online world. However, it is unclear when the wider society will reap the benefits of such a media literacy program and the health of democracy cannot wait for a younger, more media-aware, generation to grow to maturity.

Problems arising from the spread of misinformation are not confined to these online spaces. Last year, mob lynchings across India were fuelled by the spread of disinformation spreading across encrypted messaging apps. The tension between privacy and data security means that harmful content can spread like wildfire without anyone being held accountable. Since then, WhatsApp has restricted forwarding messages to a maximum of five people.

Removing or reducing the impact of content that contains verifiably false assertions is difficult but tractable. Traditionally, the role of debunking deceptive claims has fallen to competent journalists. Given the mammoth scale of the problem, algorithmic interventions are needed. However, outsourcing the ‘half-truth problem’ solely to algorithms has thus far proven ineffective and exceptionally difficult to handle, in part because cases of misinformation are often not clear cut and rely on careful interpretation.

Given these difficulties, the case for public service organisations like the ³ÉÈËÂÛ̳ to take a leading role in the fight against online misinformation is a strong one. Widespread online misinformation strikes at the heart of our public purpose to provide accurate and impartial news. However, the size of the challenge is unprecedented. Our online information ecosystem is polluted.

For its part, the ³ÉÈËÂÛ̳ is committed to being part of the push back against the wave of misinformation, distraction and deceit that characterises parts of the online world. Over the coming months, the ³ÉÈËÂÛ̳, alongside other organisations, will be looking at a whole raft of practical actions that might be taken to address misinformation across the media landscape. These interventions will sit alongside our continuing editorial coverage and initiatives like the ‘Beyond Fake News’ project.

Our approach will be cross-disciplinary; connecting researchers, designers, academics, policy makers and technologists with journalists. The impact of misinformation reaches far and wide. This conversation is not just about journalism; it’s about preserving the information that underpins society, it’s about policy, technology, humanitarian organisations and public trust.

Neither the ³ÉÈËÂÛ̳ nor its partners will entirely solve the problem of misinformation, online or offline, but we are doing our part to ensure that trustworthy information derived from competent, honest and reliable sources continues to flow freely across society, giving audiences around the world a space where they can find news reports they can rely on.

On Wednesday May 1st, I attended the latest edition of the ³ÉÈËÂÛ̳ Machine Learning Fireside Chat series on predictions and the human centric creative process.

“This time we will explore the state of the art of collaboration between machine generated predictions and people in the creative process. What is the work ahead of us in order to make this joint venture better? What is going to be the role of increased interpretability of predictions, the role of interfaces and how the work of content creator is going to change? Finally what are the dangers, and consequences to the business and the audience?”

The panel hosted by Magda Piatkowska, Head of Data Science Solutions at ³ÉÈËÂÛ̳ News, included Greg Detre, Chief Data Scientist at Channel 4, and Atte Jääskeläinen, a renowned ML researcher and former Director of News and Current Affairs at Finnish station YLE, considered to be a leading innovator among public service broadcasters.

The panel kicked off with a direct question. Are we fearful about how machines will destroy creativity and replace journalists?

Atte shared his view that there is indeed fear. Different organisations have different communications and leadership attitudes, which influence whether the landscape is presented as a threat or an opportunity. The nature of work will change and we need to question how creative work can be more productive, interesting and creative. Greg pointed to Aristotle and 19th century Luddites to articulate the existence of a long term considerable fear. 

ML Tech as Enabler or Disruptor?

Magda probed which stages of the newsroom process, from gathering to publishing, are most likely to be impacted. Atte pointed to the business models funding journalism, specifically examples of Facebook and Google who are targeting users in better ways than traditional news organisations. Greg expected change to take a long time, and reminded us that Deep Blue vs Kasparov happened in 1997, over twenty years ago.

Humans are in the driving seat, adding something to machine output. What will the interfaces look like? Greg expected this to be more like dialogue. At Channel 4, human experts use a combination of historical data and research for forecasting. By replacing the human process with machines, it will sometimes be better and sometimes worse. Humans understand warning signs and can interrogate and tweak accordingly. Atte noted that with media, people pay for negative feelings. Magda added that Public Service Media should remember that private media consider public interest as part of the optimisation of algorithms too.

Collaboration: Human and Machine?

Atte provided two examples. Firstly, the prediction of successful news stories; and second, for commercial news organisations, the prediction of how a digital news story contributes to a willingness to buy a digital subscription. But we need to manage expectations. The machine may not be able to predict as well as the newsroom thinks. Most news organisations are trying to find ways of being more interesting to younger audiences but we have an issue with identifying how we engage with younger people. Greg spoke of empowering editorial teams. Algorithmic recommendations are based on original ideas of human beings. So how do we make the machine’s response better? How do we build trust so that the machine will continuously do well for the audience?

We acknowledge that failure is not catastrophic in the context of TV recommendations, not like it could be in the case of AI driven cars. Atte contended that we need to make it clear that things need to change and to create a positive environment where change is seen as a positive opportunity. News editors can be afforded more time for thinking. This can be a positive disruption. People can concentrate on doing things they are more interested in and in being more creative. Greg pointed out that hardware is evolving so much slower than software. He likened the current climate to a dancing bear. The bear may not dance well but there is general amazement that the bear can dance at all. If  is good enough to generate fake news, it is good enough to be dangerous.

An audience member echoed the fact that most algorithms are goal seeking. Is there a danger that AI journalism will optimise for what is NOT best for society? Atte agreed that there is a danger. But stated that we must move forward and design good algorithms. Greg took the example of recommendations. What would someone who had watched everything and who knew you well recommend to you? Look at how an algorithm does this. It is possible to layer on a human component to interject.

Atte shared a tale of AI whisky. There are an infinite number of possibilities for raw whisky in barrels. The algorithm generates the ideas. However, to fit with the brand of the whisky house, there are limited proposals. So the final selection decision is made by the master blender. This first AI whisky is due to hit the shops in September! Atte’s example is about getting to the root of what is valuable.

Greg shared his definition of intelligence as “flexible, gold rated behaviour”. It is that flexibility that makes us interesting as human beings. The quality of journalism is a multidimensional issue according to Atte.  Algorithms don’t grasp contextual understanding of what words mean, like humans do. Atte however maintained that algorithmic bias is easier to understand than human bias in a way. ML can help the newsroom in the way they think by providing a mirror. 

From a commercial perspective, cheaper to produce is not always the answer. The real difference comes from where the real value is created, by a very small portion of the production. Today, as we measure consumption, we notice that only a small portion of articles are consumed. Do we want better content or more content? Greg is sceptical whether AI has anything to contribute here. We need to consider focus; where we can make an impact rather than trying to do everything. Atte added that it is hard to compete with the giants with large amounts of data and the costs of competing are high. Products being developed for other industries can be applied in media however.

Audience Discussion

Greg told us that it is possible to craft images that can trick the algorithm into thinking it is something else. It exists and will get more subtle and we should be worried about this. Atte said this reminds him of when people didn't know what bots were. The fake news problem is not about ‘do you trust this piece of news?’ but rather ‘do you trust this news organisation?’.

How do we measure trust and reflect that in how we build it into the system? An audience member stated that looking for a particular journalist is in actual fact searching for a particular bias, and so this is no different to ML in his view. Atte’s take is a little different as he believes it’s down to whether you share the values of those who tell you the stories. In the UK there are a number of political views shared through the newspapers. At Channel 4, everything consequential is checked by humans. Greg added that he thinks human concern will fade when machines become increasingly ‘right’. But they will still sometimes get it wrong. There is a problem of societal trust in media in parts of the world where media isn’t trustworthy.

The criteria for a successful news story?

There are more and more ways to engage people to pay a subscription and to stay as subscribers. Atte questioned “what should be the indicators of success?”. The answer is in the the value that content is creating for our private life and for society. The ³ÉÈËÂÛ̳ is working heavily in this area. The European Broadcasting Union will also be focusing more on this next year.

Madga sees it as a numeric and simple equation, whereby success is increased page views shared across news sites. Content, for example, that stimulates public discourse such as the environmental conversation from Blue Planet. This was a story of huge impact story prompted both public conversation and action. However this is anecdotal and difficult to measure. ‘Impact’ as a metric is a difficult one.  One person’s positive can be another person’s negative and if you define impact yourself, it is your measure. Atte added that monetary bonuses negatively impact the quality of work, thinking and creativity as we can’t know in advance what the most successful solution will be.

A human connection to the audience must be maintained. People are appreciating better selections of content. A solid voice can only be created by humans working alongside machines. Atte maintained that organisations will have to specialise more, to take care of one repertoire of the audience and to be the best. When the machine is right and we don’t take the recommendation, we overlook the better decision.

Greg picked up on how our relationship as designers of complex adaptive systems will change. We think of ourselves as controllers / programmers but really the role is more one of guidance, like an orchestra conductor. The relationship will need to evolve; our relationships with machines will change. Atte closed on the point that we tend to reject ideas that go against our human biases. But machines create ideas from which we can choose. There is a cultural change required in thinking that someone can have a better idea than you.

How easy is it, really, to achieve social good with AI? The decidedly thought-provoking discussion at the recent '³ÉÈËÂÛ̳ Machine Learning Fireside Chats presents: AI’s Hurdles to Doing Good’ was hosted by Tse Yin Lee, a senior journalist in News and Current Affairs.

The provocation...
We’ve seen AI’s use in the commercial sector explode, but what about the charity sector? There is huge room for AI to do social good, but given that charities are not traditionally at the sharp end of technology, the path is not without challenges. Is the data available to support the effort? Are there enough experts who understand not just algorithms, but the environments they’re working in? How do you even begin to check your project for ethical risks? How do smaller charities cope? How do you compete with larger companies who can pay staff significantly more?

The panel...

Julie Dodd, Director of Digital Transformation and Communication @ Parkinsons UK
Julie is Parkinson’s UK’s director of digital transformation and communications. She’s also written The New Reality, a study into how non-profit organisations can approach digital transformation and use digital technology for change. Among other things, Parkinson’s is currently working with Benevolent AI to try and identify at least three currently available medicines that can be re-purposed to address the condition and two brand new ways to treat it with new drugs.

Giselle Cory, Executive Director @ DataKind UK
DataKind is an organisation that is itself a charity whose volunteers support other charities and social enterprises, providing data science expertise and services such as exploratory analysis and prototyping. Projects they’ve worked on include helping homeless organisations identify issues people sought advice on before becoming homeless and analysing possible reasons why people progress differently in a shelter. It’s helped Global Witness to uncover potential corruption with open data and the Citizens Advice Bureau to better identify emerging social issues to support their staff training.

Michelle Eaton, Programme Manager in the Government Innovation Team - Data @ Nesta
Michelle worked with the Essex Police Force for 11 years - starting in community policing before moving on to performance analysis, and then to strategic change, where she led the Essex Centre for Data Analytics. Now she oversees Nesta’s Offices of Data Analytics programme, which helps governments reform their services, and recently wrote a blog defending the use of AI in policing.

The discourse…

Giving the panel a chance to show off, Tse posed her first question: “What’s the most exciting thing you do with AI?”

Giselle jumped in first and explained how DataKind had worked on a predictive model for a food bank to help identify people in the line of the food bank who might be in need of extra support. Exploring the topic more, Giselle was clear in stating that this model, and others like it, should exist to help advisors, and not act as the primary decision maker; relying wholly on AI-generated decisions is a dangerous game to play, especially with decisions that have deep ethical grounding.

Julie followed, explaining how they, Parkinson’s UK, have partnered with Benevolent AI to aid with the complex process of drug discovery. Their platform impressively trawls through millions of Parkinson's-related papers, looking for clues which humans may have missed. They hope to shorten the process in finding answers for new treatments to Parkinson’s.

Michelle explained how, at Nesta, they are not immediately looking to implement AI solutions to solve problems; instead, they are more focused on looking into the wider problems, from problem identification through to implementing a solution, where AI is not always the answer.

Tse then went on to ask the panel about perceptions from leadership of AI in organisations, from before the the introduction of AI to the organisation, through to implementing a fully AI-driven solution.

Continuing her earlier train of thought, Michelle highlighted how “an AI solution is not a silver bullet”. She went on to explain that a crucial part of the process is gathering data and understanding the problem space, and communicating these findings back to leadership. Only then, once you have your findings and have formed good partnerships, do you progress forward. Luckily, the leadership at Nesta was very receptive to this process.

Giselle talked next, paraphrasing the quote from , “If it’s written in Python, it’s probably machine learning, if it’s written in PowerPoint, it’s probably AI”. She went on to explain that “AI as a concept is hard to hold”, and there’s lots of steps involved in getting actual AI into non-profits; there is a spectrum of ‘data maturity’ and most small non-profits are clustered towards the bottom - which is where DataKind come in to help. A lot of what they do is completely unrelated to Data Science, it’s about holding hands, creating a safe space in order to empower people to feel confident when using data. Julie explained how, in a lot of cases, for charitable organisations the culture is somewhat ready to embrace AI, but the data is not accessible. Where she has started to see success is when the focus is put on gathering the right data first, and then using it to effectively present the business case to leadership so the projects have full backing and understanding throughout the organisation. In addition, when taking on a new project, they set up a periodic review to asses ‘data maturity’ across the organisation. She added the amusing observation, made through the questionnaires, that as the staff understand the problem space more, the scores for data confidence go up, and the scores for data accessibility go down!

“How easy is it to recruit the data scientists you need?” Tse looked to the panel for answers.

Charities often can’t afford to pay market rates. Julie explained how at Parkinson's UK, even though they have over 500 people, that’s not enough - they need more to scale. The job market is currently squeezed for data roles, causing market rates to increase beyond what charities can afford to pay; they find they get candidates who either join for the love of the cause, or candidates who enjoy the flexibility that comes with working in a charity. Drawing on her experience as a police community support officer, Michelle explained how she traversed through various roles within the police and eventually came out in a rather different role, overseeing Nesta’s Data Analytics programme, and its thanks to this journey that she greatly values aptitude and attitude over skills.

Tse then talked through a list of somewhat provocative headlines negatively covering AI in healthcare and policing and asked the panel what they thought of them.

Acknowledging her bias, Michelle began by stating she’s emotionally invested in the police so it’s difficult to criticise them. Due to factors such as the police being stretched, demand rising, and the front line shrinking, the police have to look for new ways of doing things, complementary methods. These stories often miss the outcomes of projects, they miss the actual good. The positive aspects, such as doing these things to better understand threats to the public, and understanding how to protect people better, are rarely talked about. She then continued discussing how the police are making a real effort to develop these capabilities transparently and are considering all the ethical implications upfront, whilst in constant communication with experts. Julie made an interesting point that the negative headlines in healthcare have acted as the seed for more ‘grown up’ conversations about AI in healthcare - she says it grounds these concepts to reality and moves us away from the ‘AI will save us all’ and ‘AI sounds great, let’s do a thing’ mindsets.

“Do you think the sense of mission in a charity can blind them to the possible ethic problems of the way they are using data?”, Tse asked.

Any organisation dealing with vulnerable people tend to approach their cohort with a duty of care and are therefore generally cautious with these kinds of things, Giselle explained. That doesn’t mean there won’t be mistakes, she continued, there are examples of partnerships between charities and commercial sector companies which may not be entirely ethically sound. Expanding on partnerships, Michelle talked about Parkinson’s UK’s partnership with Benevolent AI and the challenges they have faced with data ownership, intellectual property, and the immaturity of legal frameworks surrounding partnerships such as these. She also expanded on the difficulties of reward - should the person's data who it was originally from be rewarded? Julie added that with charities, strong ethical frameworks are usually fundamental to the organisation so there’s less concern about misuse of data compared to the commercial sector.

The floor was then opened for questions from the crowd. We discussed Geographical differences in data ability, Giselle discussed how some areas have very strong government specialists, some have lots of highly skilled people volunteering, and some areas really struggle for non-profits. A question on using data sources in charities as a revenue generating asset was posed to the panel; Julie informed us that charities in the UK are often working together and sharing a lot, and there are rich ongoing conversations regarding how to make open data platforms; Michelle added that Nesta are looking into data trusts and data collaboratives as a way to incentivise sharing.

After a few more rounds of questions, some lovely nibbles, and a beer or two, the event drew to a close!

For more details about upcoming events, visit .