I recently organised an AI Safety x Content Fellowship for Africans through CoLoG Africa.
The idea was to identify people interested in communicating AI safety, equip them with the skills and tools to produce accessible content, and use content creation to expand awareness of AI safety across African communities.
I went into the programme with a number of assumptions about what participants already knew, the tools they would need, and what would make the programme successful.
Some of those assumptions were wrong.
Here are some lessons I learnt.
1. A computer science background does not automatically translate into AI or AI safety literacy
One of the more surprising things I learnt was that having a computer science background does not necessarily mean someone understands AI, particularly the emerging field of AI safety.
Approximately 80% of the fellows had a computer science background, either at undergraduate or postgraduate level. I therefore expected most participants to have a reasonable understanding of artificial intelligence and some familiarity with the conversations happening around AI safety.
I was wrong.
Many participants were relatively unfamiliar with important AI safety concepts, including alignment, interpretability and other areas of technical AI safety research.
This was an interesting observation because it challenged my assumption about what a computer science education prepares people for.
Studying computer science does not automatically expose someone to frontier AI research, the technical challenges of alignment or the broader questions surrounding increasingly capable AI systems.
I had to introduce additional workshops to bridge this knowledge gap.
In retrospect, this should have been part of the original programme design. We needed to establish a common foundational understanding before expecting fellows to communicate more advanced AI safety concepts.
Lesson: Do not use academic qualifications or technical backgrounds as proxies for AI safety knowledge. Even when recruiting computer science graduates, build foundational AI and AI safety literacy into the programme.
2. Distribution over Technical expertise
Another interesting observation was the difference in how participants approached content creation.
I initially thought people with technical backgrounds would be particularly enthusiastic about producing videos explaining AI safety concepts.
Interestingly, they were less inclined to share their work publicly, particularly on social media.
Students and early-career graduates, on the other hand, were much more enthusiastic about publishing their videos and sharing their work with their networks.
On the other hand, the fellowship had two(2) journalist in the cohort. Once the journalists understood the subject matter, they were able to translate what they learnt into stories that reached audiences far beyond our fellowship. Some of these stories were aired on their respective national television stations.
Think about the difference in distribution. A fellow publishing a short video on LinkedIn might reach a few hundred people. A journalist producing a television story can potentially reach thousands or millions.
This is not to diminish the value of individual content creators. Rather, it highlights the importance of identifying participants who already have access to established distribution channels.
Lesson: Future cohorts will prioritise journalists and media professionals. If the objective is scalable public awareness, distribution networks should be considered alongside individual content creation skills.
3. AI video production is more expensive and technically demanding than it looks
I underestimated how difficult it would be to produce good AI-generated videos.
There is a tendency to assume that AI video platforms have made content production straightforward. You provide a prompt, generate a video and publish it. In practice, producing something genuinely good requires considerably more work.
First, the platforms are more expensive than I initially anticipated. Access to the better tools often comes with subscription costs and usage limitations.
Second, there is a significant learning curve.
Getting good results requires understanding how different tools work together. You might need one platform for script development, another for generating visuals, another for editing and another for voice production.
Even generating a good script requires more than a simple prompt. You need to understand storytelling, visual direction, scene transitions and how to translate abstract concepts into something people can understand.
I also discovered that reference videos matter considerably. Having good examples of the type of video you want to produce makes it easier to guide the creative process.
One particularly useful discovery was voice cloning. Using a cloned version of your own voice as the background narration often produced better results than relying on generic AI-generated voices. The latter can sound artificial and make an otherwise good video feel less authentic.
Lesson: AI video production should be treated as a skill, not simply as access to a tool. Budget for subscriptions, experimentation and the time required to learn the platforms.
4. Tool orientation should happen before production begins
Closely related to the previous point is the importance of expert-led workshops.
It is not enough to give fellows access to AI video platforms and expect them to figure everything out.
Even when platforms appear intuitive, participants need to understand their features, limitations and practical applications.
I would recommend introducing dedicated expert workshops early in the fellowship.
These workshops should take participants through the entire production process, from developing a script to generating visuals, producing voiceovers and editing the final video.
The objective should be to reduce the time participants spend figuring out the tools and increase the time they spend creating.
This is particularly important when working with people who have different levels of technical experience.
Lesson: Bring in practitioners who already understand the tools. A two-hour practical workshop could save fellows several days of experimentation and money.
5. Real-world AI stories are powerful learning tools
One thing I noticed was that sharing AI news and developments I came across on X (formerly Twitter) generated considerable interest among fellows.
Stories about new AI capabilities, emerging risks, model releases, AI-related incidents and developments in the industry often sparked more curiosity than abstract discussions about AI safety.
I think this is an important lesson about how we introduce AI safety to people who are relatively new to the field.
It is easier to get someone interested in a concept when they can connect it to something happening in the real world.
For example, discussing prompt injection becomes more interesting when accompanied by a recent incident demonstrating how it can happen. Similarly, discussions about frontier AI capabilities become more engaging when participants can see what the latest models are capable of doing.
I would encourage facilitators to actively bring current AI stories into sessions rather than relying exclusively on prepared learning materials.
Lesson: Use real-world AI developments as entry points into AI safety concepts. Stories create curiosity, provide context and make abstract risks easier to understand.
6. Smaller breakout rooms can improve engagement and preparation
Another thing I would change is how we structure group discussions.
I think facilitators should introduce more discussion-based sessions using smaller (not more than 3) breakout rooms.
Rather than allowing participants to choose their own groups every time, I would recommend randomly assigning fellows to different pairs or small groups during each session.
The idea is to ensure that participants interact with different people throughout the fellowship rather than forming familiar groups and sticking to them.
This has several benefits.
First, it encourages participants to prepare for discussions because they know they will be expected to contribute to a smaller group.
Second, it creates opportunities for fellows to get to know one another beyond the formal fellowship activities.
Third, it exposes participants to different perspectives and ways of thinking.
For a remote fellowship, where participants may never meet physically, creating these opportunities for interaction is particularly important.
Lesson: Make breakout discussions a deliberate part of the fellowship design. Rotate participants randomly across sessions to encourage preparation, participation and relationship-building.
7. Psychological safety matters, even in a content fellowship
One observation I found particularly interesting was how participants behaved during discussions.
Participants tended to be more expressive when they were not on live video.
There was a noticeable difference between conversations where people were expected to appear on camera and those where they could contribute without that pressure.
This is worth thinking about when designing remote fellowships.
We sometimes assume that video calls automatically create better engagement. But being visible on camera can introduce a degree of self-consciousness that affects how freely people contribute.
People may have ideas but hesitate to express them because they are concerned about how they sound, look or might be perceived.
I would therefore place greater emphasis on creating spaces where participants can contribute comfortably, including smaller discussions, audio-based conversations and written reflections.
Lesson: Participation should not be measured by how comfortable someone is appearing on camera. Design multiple ways for people to contribute.
8. Mentor feedback works better in a huddle than in document comments
Another practical lesson concerned mentorship.
Initially, I expected mentors to provide feedback through comments on shared documents.
While this is useful for correcting specific issues, I found that it was not always the most effective way to help fellows improve their work.
Feedback sessions conducted through smaller group discussions or huddles worked better.
There is something valuable about being able to discuss a script, explain why a particular approach might not work, exchange ideas and collectively identify improvements.
A document comment can tell someone what is wrong. A conversation can help them understand why it is wrong and how to approach the problem differently.
Huddles also create opportunities for fellows to learn from feedback given to other participants. This makes mentorship more efficient because the learning is not restricted to the person receiving the feedback.
Lesson: Use document comments for specific edits, but make regular, interactive mentor huddles the primary feedback mechanism.
9. Fundraising for AI safety communication programmes is harder
Perhaps one of the most difficult aspects of organising this fellowship was raising money to support it.
Eventually, I had to use my personal savings to support the fellowship.
This was not necessarily how I had planned to finance the programme, but I was committed to seeing it through.
One thing I learnt is that funding for AI safety communication, particularly programmes targeting African audiences, is not always readily available. There is funding for technical research, policy work and other established AI safety activities, but translating the importance of communication and public awareness into a fundable proposal can be challenging.
I also underestimated how much money would be required to deliver the programme properly, particularly when accounting for AI video subscriptions, production credits, mentorship and other operational expenses.
Lesson: Start fundraising well before launching the fellowship (6months ahead).
