I’m back from London.
It was as beautiful and alive as ever, but also hot as hell.
There was a record-breaking heatwave while I was there, and the city handled it with its usual strategy: open windows, cold pints, heroic little fans, and the sense that air conditioning would be lovely, but perhaps a bit much.
Still, I loved it.
I had some good meetings, saw friends, walked as much as I could without melting, and remembered how much I love being in a city that feels that awake.
My sister was there for work too, though she ended up working about 95% of the time.
So our quality time mostly happened in tiny stolen pockets, which we tried our best to document.
Anyway. Back to business.
Every so often, an AI study makes me read the same sentence three times and then say, “Wait… what?”
This was one of those.
Researchers at the University of Oxford and the UK’s AI Security Institute set out to test whether today’s most advanced AI systems could out-persuade people who are genuinely good at changing minds.
Not random people.
Top performers from an online persuasion tournament. Professional fundraisers. Elite competitive debaters, including world champions.
The humans did not walk in cold.
The debaters chose the issues they thought they could argue best. They researched in advance. They practiced. They competed for £1,000 bonuses.
AI still won.
Across four experiments and nearly 19,000 conversations, the AI systems were more persuasive than every human group tested.
So the researchers tried to help the humans catch up.
They built the debaters a coaching tool that let them practice against the same AI that beat them, see the instructions it had been given, study their own past conversations, and compare their replies with what the AI would have said instead.
AI still won.
Then came the part that moved this from interesting to unsettling.
Participants were given a small amount of real money and the choice to keep it or donate some of it to Save the Children.
Up against trained fundraisers—from a firm that had spent seven years raising money for that exact charity—the AI raised nearly three times more.
It was a tiny amount of money, so let’s not overclaim it.
But it was still a real choice.
So yes, AI can persuade.
And it can move people to take action.
But How Did It Win?
When I first read the study’s headline, I assumed the AI might have won because it was better at the human-feeling parts of persuasion.
More emotionally intelligent. Better at understanding psychology and making people feel seen and understood.
But this study’s findings point somewhere else.
AI could deliver more relevant information, faster.
The elite debaters wrote about 54 words per reply and took roughly 95 seconds to respond.
The AI wrote much longer tailored replies almost instantly.
When researchers forced it to slow down and respond at human speed, with human-length messages, its advantage disappeared.
Obviously, emotional resonance, rapport, and trust are still central to persuasion.
But AI’s edge here came from a combination humans struggle to match: it can feel responsive and attuned while delivering credible-sounding information and relevant arguments faster than anyone can fully evaluate them in the moment—and it can keep doing that without ever getting tired, defensive, or bored.
Which is exactly the kind of thing AI systems can scale infinitely.
A human fundraiser has only so many good conversations in a day.
A political volunteer can only knock on so many doors.
A strategist can only test so many arguments.
AI removes a lot of that friction.
Persuasive, personalized, one-on-one conversation becomes cheaper, faster, and easily repeatable.
Some of that can be genuinely useful.
A health group could help people understand vaccine safety or quit smoking. A small nonprofit could raise more money without a giant fundraising team. A legal aid group could help people understand their rights. A teacher could help a student see a hard topic from a new angle.
But the same capability is available to anyone with a goal and a budget: companies trying to sell more of their stuff, campaigns trying to move voters, platforms trying to hold attention, bad actors trying to spread misinformation, and scammers trying to steal your money.
And we’re nowhere near ready for this.
Persuasive AI will not wait for society to agree on norms.
It will be used by whoever sees value in changing minds.
Which means the real questions are what we allow it to be used for, what goals it is optimized to achieve, and who benefits when it succeeds
📈 Work with Me
AI Advising and Consulting: Strategic guidance to build sustainable AI strategies and adapt to shifting audience behavior
AI Training: Practical training that helps teams use AI more strategically, effectively, and safely—so it becomes a trusted thought partner, not just a task assistant
What You Need to Know About AI This Week ⚡
Clickable links appear underlined in emails and in orange in the Substack app.
⚡ ChatGPT had a very big week.
1️⃣ OpenAI’s answer to Claude Fable is here.
OpenAI launched GPT-5.6, its long-awaited new model family.
The one to pay attention to is Sol, the most powerful version.
It’s clearly being positioned as OpenAI’s answer to Claude Fable 5. On OpenAI’s benchmark for long-running professional workflows, Sol beat Fable 5 while coming in at a much lower estimated cost.
So yes, this looks like a very big release.
But it launched yesterday, and I need a week or so to test it before I can share anything useful about how it performs, how it compares to Fable, and how to get the best results from it.
I also expect many of our prompts will need adjusting, but it’ll take some experimentation before I can figure out what changes actually help in practice.
OpenAI also launched ChatGPT Work, a new agent inside ChatGPT that can pull context from your tools and files, plan the approach, and create things like docs, decks, spreadsheets, analyses, dashboards, websites, and web apps.
That should be a big deal.
But honestly, everything about this launch was so confusing and overwhelming. Even for me. And I know my way around this app.
ChatGPT Work seems like a version of Codex for non-technical work, but I need to spend more time with it before I try to explain what it actually is, how it’s different from Codex (likely just doesn’t show the code behind the executions), and where it’s most useful.
And this will take me a while…
2️⃣ ChatGPT Voice just got a major upgrade.
OpenAI also launched GPT-Live, a more advanced and smarter ChatGPT Voice model that can listen and speak at the same time.
That changes the feel of the interaction.
Instead of waiting for clean turns, GPT-Live can keep listening while it responds, handle interruptions more naturally, pause when you need a second, use web search and memory, show visual results, and move between voice, text, and images in the same conversation.
OpenAI says harder questions that require in-depth search or advanced reasoning are handed off to a more powerful frontier model in the background while the voice conversation keeps moving.
If you’re on a paid plan, you can set voice mode to a higher intelligence level. Go to Settings → Voice → Intelligence and set it to High.
That’s just my preference. You may wait a few extra seconds (or more) when it needs to search, reason through, or analyze something, but I’d rather wait for the smarter answer.
I’ve only had two days with this smarter voice mode, but I’m already in love.
On a walk yesterday, I spent about ten minutes talking through last weekend’s box office numbers.
It pulled exact grosses, compared domestic and international performance, answered follow-up questions by market, looked at audience scores, and helped me think through what new releases are tracking at for this weekend.
ChatGPT’s smartest models could already help with this kind of analysis.
The difference is that now I can do this kind of work while walking around or driving.
I know a lot of you are insanely busy with work, family, and other life stuff. But you need to unbusy yourself long enough to watch OpenAI’s 18-minute livestream 👇.
Why? Because this is how a “cool demo” becomes a “wait, what happens to my business?” moment.
A thing that used to require a separate app, tool, vendor, workflow, or entire team becomes something you can just do by talking to ChatGPT.
For some companies, that makes the thing they sell far easier to replace.
For others, it lets them add capabilities that make what they already offer much more valuable.
⏳ Fable is back for a few more days. Here’s how I’d use it.
The model launched, got pulled after U.S. government export controls, then came back online with stricter safeguards.
A lot of drama for one model.
Anthropic has now extended Fable access for paid Claude subscribers through July 12, before the model moves to usage credits.
So I’ll be spending most of my weekend continuing to test it while I can.
I’m still trying to wrap my head around this model, and how and when to use it.
When a model has genuinely new capabilities, a few experiments don’t tell you enough.
You need reps. But every turn burns through your allowance, and the usage caps are tight, so you have to be much more intentional about what you test.
So far, my suggestion is to use Fable for complex, multi-step projects and well-specified problems.
It seems especially useful for end-to-end work that would take a person hours, days, or even weeks to complete: planning, research, analysis, workflow design, coding (obviously), and projects where the output is concrete enough to verify because there are right and wrong answers.
I’ve also had some strong results using it for more open-ended strategy and brainstorming work, but that takes more steering, more context, and way more judgment from you.
The model sometimes speaks in a way that feels almost alien.
I find its language hard to follow at times, and difficult for my brain to process.
I often have to read a sentence a few times, and have even asked ChatGPT or Claude Opus 4.8 to translate it for me.
I’ve also been getting better at prompting it to speak my language, which takes some practice but helps a lot. Like a lot a lot.
But what’s becoming clear is that getting the full value out of this new class of models is a skill in itself.
👉 I’m still learning, but here are a couple of tips based on my testing so far:
Give it your most challenging problems and ambitious projects. Pick a task harder than what you’d assign to prior models, and have Fable 5 scope it, ask clarifying questions, create a plan, and then execute.
Give it the why behind the request. Fable tends to perform better when it understands the intent behind a request rather than inferring intent on its own. Provide context about the why of your project and how you plan to use the output.
Use Fable for the hard thinking, then hand off execution. You don’t always need Fable to carry the whole project. Use it to scope the problem, reason through trade-offs, set the strategy, create the plan, and define what good looks like. Then hand that plan to another capable model, like Claude Opus 4.8, Sonnet 5, or GPT-5.5, to execute the parts that don’t require Fable (you can FINALLY switch models within a chat in Claude). Don’t make it do the grunt work.
🎭 YouTube’s faceless creators are hiring faces.
As YouTube cracks down on AI slop, some faceless channels—channels built without an on-camera personality, often using narration, clips, graphics, and AI visuals—are losing revenue or getting demonetized.
So some creators are now hiring inexpensive hosts from platforms like Fiverr and Upwork to be the face of their videos.
🛒 AI shoppers are arriving ready to buy.
U.S. shoppers coming from ChatGPT, Gemini, and other AI tools generate 53% more revenue per visit, convert 54% more often, spend 53% more time browsing, and view more pages than shoppers from non-AI sources.
This means people using AI to shop are showing up with stronger intent.
They’ve already compared options, asked follow-up questions, and narrowed their choices before they ever land on the retailer’s site.
🖼️ Getty Images is coming to ChatGPT.
Getty Images and OpenAI have struck a multi-year agreement that will let Getty’s licensed photo library appear inside ChatGPT’s search and discovery experiences. Financial terms weren’t disclosed, but investors liked what they saw: Getty’s stock initially surged after the news and was later trading at more than double its pre-announcement price.
The deal comes while Getty is still fighting AI companies over the use of copyrighted images as training data.
This agreement gives OpenAI a way to show high-quality, licensed visuals in ChatGPT, and gives Getty a new way to make money from its image library.
The catch: a licensed photo can still contain rights Getty may not fully control, like a celebrity’s likeness or a brand trademark. So ChatGPT showing a Getty image does not automatically make it safe to reuse in an ad, product, campaign, or other commercial context.
🎬 A24 got Google’s $75 million—and an AI backlash.
A24 and Google have struck a multi-year AI research partnership to develop new filmmaking tools and workflows, with Google reportedly investing about $75 million in the independent studio.
The deal reportedly does not give Google access to A24’s film library, TV library, or data. A24 gets early access to DeepMind’s research and infrastructure. Google gets feedback from filmmakers and production teams who can help make the tools more useful in the real creative process.
But the backlash is showing up on the A24 subreddit, where some fans are posting that they’ve canceled AAA24, the studio’s paid membership club.
For a studio whose brand has been built on artists, taste, and creative credibility, even a carefully scoped AI partnership can feel like a breach of fan trust.
For creative companies, AI announcements now carry an emotional charge and need to be handled with lots of care.
🧑💻 AI is starting to do work people pay humans to do.
A new Remote Labor Index study tests whether AI agents can finish real online freelance projects end to end—the kinds of jobs someone might hire out, like making a short animated ad, designing product visuals, analyzing data, building a web app, or turning photos and measurements into a room redesign plan.
AI still fails most of the time, and progress is uneven. But AI can now complete some valuable digital tasks while still stumbling on things humans might find easy to do.
The main story is the pace of improvement: the top model success rate jumped from 2.5% to 16.1% in under eight months, with Anthropic’s Fable 5 currently leading the benchmark.
Job disruption may show up less as entire roles changing overnight and more as paid work getting unbundled.
AI handles more of the doing, while humans shift toward deciding what the AI should do, briefing it, supervising the process, evaluating the output, and deciding what’s good enough.
🏛️ OpenAI’s next big pitch: give the public a stake in AI.
OpenAI has reportedly proposed giving the U.S. government a 5% stake in the company, worth about $42.6 billion at its recent $852 billion valuation.
The idea would extend beyond OpenAI, with Washington holding similar stakes in other leading U.S. AI companies—though it’s not clear those companies would agree.
The public-facing argument is that ordinary Americans should share in AI’s upside, especially if the technology reshapes jobs, wealth, and power.
President Trump has said he’s exploring ways to give Americans a stake in leading AI companies, while Bernie Sanders has pushed a much more aggressive 50% public ownership plan.
This also looks like a bid for political legitimacy.
AI still feels to many people like something happening to them, not for them. A public stake could help OpenAI argue that the benefits won’t flow only to founders, investors, employees, and tech giants.
🤔 Why big AI labs are hiring so many philosophers.
Top AI labs are hiring philosophers because models now need rules for messy situations: when to push back, when to admit uncertainty, how to avoid flattery, whose values to follow, and how to weigh harm, truth, and user intent.
Every chatbot is being taught a worldview. But whose?
In case you missed the last edition, you can find it 👇:
🤓 Claude's Most Powerful Model Is Here, and You Better Use It Soon
On Tuesday, Anthropic launched Claude Fable 5, which it says is the most capable model it has ever made available to the public.
That's all for this edition. See you next time.
Thoughts, feedback and questions are always welcome and much appreciated. Shoot me a note at avi@joinsavvyavi.com.
Stay curious,
Avi