The Perception Gap

By Ben Norman @ 2025-08-18T10:51 (+5)

This is a linkpost to https://futuresonder.substack.com/p/the-perception-gap

AI isn't slowing down, but disappointing releases make people think it is. What can be done about this?

Introduction

Previously, I raised the point that an AI winter may not actually be good for AI safety. While there are no signs of an AI winter being right around the corner (and, no, GPT-5’s release does not mean a slow-down in capabilities), the general narrative may be starting to perceive AI progress as potentially stalling. Concerningly, this has affected the views of key actors — like the U.S “AI czar” David Sacks. As noted in The Transformer:

“The biggest fallout from GPT-5’s botched launch may end up being people like Sacks having a false sense of security.”

The article also (accurately) claims the perceived slow-down has had undeniable effects on AI policymaking:

“It’s fair to say that the conversation about AI regulation and preparing for AGI would be vastly different if not for the release of increasingly impressive models in GPT-4, o1, and o3.

But GPT-5’s botched release might have undone much of that work.”

The Problem of Lacklustre Releases

I doubt GPT-5’s release is the only time we will have to deal with this issue. On the path to AGI, there will surely be other lacklustre releases. Many will use these as ammunition to support their claim that it’s “all hype”. As AI capabilities rise and the stakes get higher, the consequences of such behaviour will be more and more far-reaching. You could argue that the lacklustre releases won’t matter, as they’ll be overshadowed by ones that viscerally demonstrate high capabilities—causing policymakers to feel the AGI again.

However, this seems like an unsustainable model for awakening governments and the public to AI risk. If their views are constantly shifting every time perceived capabilities change, they will not be able to understand nuance and plan accordingly.

t’s also hard to predict what exactly may elicit strong reactions. ChatGPT acted as a fairly effective[1] warning shot, but the new series of reasoning models arguably didn’t. Why? My guess would be that you need a higher-level of understanding (of either AI progress or a scientific domain you’re an expert in) to be impressed by such models. An everyday person using ChatGPT for the first time would be blown away – as they would have never seen or had access to such capabilities before. However, now that everyone—even your grandmother[2]—has been exposed to ChatGPT, reasoning models seem only marginally impressive. You’d only understand the capabilities leap if you had a STEM background and used it to help solve a problem in your work, or if you already kept up with AI.

How Do We Fix This?

So, what can be done about all this? How can we make sure policymakers (and the public) aren’t updating too strongly on lacklustre releases? Here are some approaches that could help:

  1. Continue Sam Altman’s strategy of iterative deployment. This does seem like a good plan (and has worked) with current capability levels, but may become less ideal as models pose more risks (e.g. in helping rogue actors create biological weapons) or become too linked with national security to release to the public.
  2. Create more visceral and narrative explanations of AI risk (e.g., AI 2027, the impressive new video version of AI 2027, How Takeover Might Happen in Two Years, etc). AI 2027 has undeniably had far reaching effects on public perception/policy (even JD Vance has read it!), but it is probably still too early to measure outcomes in terms of this strategy.
  3. Create and distribute demonstrations that people can interact with. Given the effects of ChatGPT’s release, this is very important (and seemingly neglected within AI safety?) CivAI seems to be doing some very cool work in this area:

They’re also mentoring researchers in the Fall 2025 SPAR round, focusing on building “software demos that give people a gut-level understanding of AI risks.” If you think this is important and have software development skills to contribute, you should apply!

4. Mandate transparency between governments and frontier AI companies. If they are expecting a capabilities increase, policymakers should not be kept in the dark about it. This is already kind of happening – but it shouldn’t be voluntary.

The Current Landscape

To understand what's already being done in this space, I looked into current initiatives. All of the ones listed below are broadly focused on increasing the salience of AI risks to policymakers and the public — either through demos, or other means.

Conclusion

So, where does this leave us? We are arguably stuck in a cycle where every disappointing release becomes fuel for AI skeptics[3], and policymakers ping-pong between concern and complacency. This isn't sustainable, and it definitely won't work as we get closer to AGI.

I believe there are concrete things we can do about this. If you have software development skills, seriously consider building demos. We need more visceral/creative (yet well-researched and evidence-based) work to make AI risks tangible. We should continue to push for mandatory transparency between frontier labs and governments.

And (perhaps) most importantly, we need to get better at maintaining consistent narratives about AI progress. The next time someone (e.g. friends, family, etc) points to a disappointing release as proof that AGI is "all hype," we need to be ready with context about the broader trajectory. We can’t let the release cycle distract us.
 

  1. ^

    It’s arguably the case that many current AI governance initiatives (e.g., national AI strategies and ongoing international discussions) trace their beginnings back to the "ChatGPT moment” as a wakeup call.

  2. ^

    Personally, mine is even familiar with (and invested in) the drama with OpenAI’s contract with Microsoft and their AGI clause

  3. ^

    As well as those who have vested interests in denying AGI as a possibility.