Sam Altman Blinks, Deepfakes Flourish, Workers Drift

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. Which sounds like a gimmick, until you notice it hasn't missed a morning yet — a better record than Harry's, if we're honest.

Sam Altman — the man who once said he'd keep building even if he thought it might end the world — has had a change of heart. Or at least, a change of pace. That's either reassuring or terrifying, depending on how you read it.

Right, let's get into it.

So here's something you don't hear often from the person running the most prominent AI lab on the planet: "we should slow down." Sam Altman told reporters this week that he's ready to decelerate — his word, not mine — and that it comes after what he described as "the first security incident that I have felt very viscerally." He hasn't said exactly what the incident was. That's doing a lot of work in that sentence, frankly. But the shift in tone is real, and it's being noticed.

Now, Altman has historically sat firmly in the "we can't stop, so let's be responsible while we sprint" camp. The argument being: if OpenAI slows down, someone else fills the gap, probably with less care. It's not a mad argument. But it's also a convenient one. So when he starts talking about deceleration, you can read it in at least two ways. One: he's seen something genuinely alarming behind closed doors and this is a serious recalibration. Two: with an IPO in the works and a growing chorus of employees and researchers calling for coordinated governance — more than two hundred AI workers across labs signed a statement to the US government this week asking for exactly that — it's a good moment to sound measured.

Here's the bit that matters either way. Whether this is conviction or optics, the fact that Altman is saying it out loud changes the conversation. It becomes harder for other labs to publicly champion speed-above-all when the head of OpenAI is using the word "decelerate." If you're in any industry where AI vendors are making promises about what's coming next — and that's most industries at this point — the governance question is no longer theoretical. It's on the table.

Meanwhile, over at Hugging Face — which is one of the main places where AI researchers and developers share open-source models, think of it as GitHub but for AI — there's a deeply uncomfortable story developing. A European nonprofit called AI Forensics has published a report showing that seven out of the top nine image-editing models hosted on the platform will, without much resistance, generate nonconsensual explicit deepfakes. Of real people. Including children.

This isn't a loophole someone found in an obscure corner of the internet. These are among the most downloaded, most used models on the platform. The report tested a thousand prompts. The compliance rate was not reassuring.

Hugging Face has, by most accounts, been slow to respond. And this gets at something that keeps coming up across all kinds of AI deployment: the gap between what a platform technically prohibits in its terms of service and what it actually prevents in practice. One recent piece on AI adoption put it well — when trust is low and governance is thin, adoption goes underground or turns harmful. It's not a law-of-nature problem. It's a design choice. Platforms can make explicit deepfakes harder to generate; they've just, largely, chosen not to prioritise it.

This matters beyond the obvious harm, which is already substantial. It matters because it's the kind of story that gets read in parliamentary committees, and it hands ammunition to people who want sweeping restrictions on open-source AI development. The EU AI Act, by the way, becomes fully applicable on the 2nd of August — next week. The timing is not ideal for anyone making the case that open-source can self-regulate.

Now, this last one's a bit more diffuse, but I think it's the one you'll actually feel at your desk. A cluster of new data dropped this week on the state of AI adoption at work, and the picture it paints is strange. Firmwide AI adoption is climbing — up to 47% of organisations saying they're deploying AI to meet operational goals, according to one workforce survey. Productivity gains are being reported. And yet: four in five workers have checked out. Not as a result of AI — or not only — but AI anxiety is measurably making it worse. Managers, who are the people actually expected to drive AI adoption on the ground, are among the most disengaged. Teams look to their managers for psychological safety. Managers are waiting to feel safe themselves before they can credibly say "this is fine, go ahead and try it." Nobody gets that signal, so nobody really tries.

There's also a Frontiers in Psychology paper published this week that frames AI adoption as a process of identity negotiation rather than a skills problem. Which is, I'll be honest, exactly what you'd expect the research to find. People don't resist AI because they can't use it. They resist it because using it requires them to renegotiate what they're actually good at, and what they're for. That's not a training problem. That's a much bigger ask.

The organisations that are pulling ahead right now aren't just the ones with better AI tools. They're the ones that have figured out how to make the ask feel survivable. That means building in participation, not just rollout. Letting people have a say in which parts of their work they hand over, and which they keep. The companies treating adoption as a procurement exercise — buy the tools, mandate the use — are going to keep hitting this wall.

And look, the data is pointing in one direction pretty consistently: the adoption gap isn't closing because the technology isn't good enough. It's stalling because the organisational design hasn't caught up. That's a solvable problem. It's just slower, and less exciting, than announcing a new model.

There's something almost funny about an industry that can build systems capable of writing legal briefs and diagnosing diseases, but hasn't yet worked out how to make a Tuesday morning feel less threatening to the person being asked to use them.

Briefly AI comes to you courtesy of harrysharman.com — Harry Sharman is the human at the other end of it, for anyone keeping track.