HAPPY FRIDAY,
October already, which means the AI hype cycle has officially survived another summer. This week's big story is OpenAI's revenue gap, which is either a bump in the road or a fairly important data point about whether the economics of this industry hold up the way everyone assumed. Probably somewhere in the middle.
In This Week's AI Fridays:
- OpenAI's revenue comes in $20bn light
- Are we putting too much faith in AI's ability to say no?
- Anthropic draws a line on fake sources, and marketers should pay attention
- Google's workplace agent is live, with caveats
- Plus: AI agents hacking Wikipedia, safety researcher firings, and a cheaper Haiku
OpenAI's Revenue Gap Is Bigger Than the Headlines
OpenAI is now forecasting $50bn in revenue for the year, which sounds impressive until you remember it was recently signalling closer to $70bn to investors. A $20bn shortfall at this scale is not a rounding error. The gap matters, but not because OpenAI is having a hard time. It matters because most AI valuations, investment theses and boardroom AI strategies from the last two years were built on a growth curve that now looks rather optimistic.
The Highlights:
- OpenAI's 2026 revenue forecast sits at $50bn, roughly $20bn below what was recently communicated to investors
- The shortfall follows a period of aggressive expansion, significant infrastructure spend, and a string of high-profile product launches
- AI stocks fell broadly after the Financial Times reported the gap, with the selloff reflecting wider anxiety about AI monetization timelines
- The gap does not mean the company is failing, it means the growth curve is flatter than the story required
The Takeaway:
I have been saying for a while that most big things turn out to be medium-sized things, and here is a $20bn data point in that direction. The technology is real; the valuation math... not so much.
The Uncomfortable Truth About AI Refusals
MIT Technology Review is making the case that the AI safety field has been dangerously overconfident in AI's capacity to refuse harmful instructions. The argument is not that refusal mechanisms are useless; it is that they have been treated as more reliable than the evidence warrants, and safety policy has been built on that assumption. That is a meaningful distinction, and one the industry has some incentive to paper over.
The Highlights:
- Current AI systems show inconsistent refusal behaviour; models that refuse harmful prompts in testing often fail under different phrasing or context
- Safety policy and public trust have both been built in part on the assumption that refusal is a solved, reliable mechanism
- Researchers are arguing that overconfidence here has real downstream consequences for how AI gets deployed in sensitive contexts
- The piece does not argue AI safety is impossible; it argues the current baseline is weaker than advertised
The Takeaway:
The field has been treating "it usually says no" as if it means "it reliably says no," and those are not the same thing. If the guardrails are doing less work than the documentation implies, that is worth knowing before the use cases get more consequential.
AI in Marketing
Claude's New Rules Target Fake Sources Built To Sway AI Answers
Anthropic has updated Claude's usage policy to explicitly call out the practice of seeding AI answer sources with misleading content, building fake or manipulated material specifically to influence what AI systems cite and surface. If you were wondering whether the AI-era version of black-hat SEO was going to get a formal response from model providers, here it is. For marketers, the more useful read is what this signals about where the guardrails are being drawn: gaming AI answers is now officially in the same category as other high-risk manipulation tactics, which means the downside risk of trying it has gone up considerably. The better investment is in content that AI systems want to cite because it is credible and specific, not because someone gamed the sourcing.
Google's Workplace Agent Is Live, With Questions Attached
Google has launched an AI agent for the workplace that can pull from a company’s documents, calendars, email and internal data to get knowledge work done. Earlier tools couldn’t move between those sources this way, so it’s a real step forward. Some big questions are still open. What will it cost? How much work does it take to connect to the systems a company already runs? And what does the agent do when the information it’s working from is out of date or unclear? None of that comes up in a launch demo. It comes up around month six.If you’re in marketing or IT and looking at this, treat it as a pilot. Find out where it breaks before you roll it out across the company.
Import AI 475: Swarm Scaling, Biology Watermarks, and the AI Science Economy
This week's Import AI covers three things: swarm scaling, where multiple smaller models coordinate rather than one large model doing everything; Google DeepMind's work on watermarking biological research outputs; and the emerging economics of AI-driven science, where the question is no longer whether AI can contribute to research but who captures the value when it does. The swarm scaling piece has the most immediate implications for anyone watching compute costs and model architecture, as it suggests the efficiency gains from coordination may be more accessible than the gains from raw scale. For marketers, the same dynamics playing out in research are going to show up in creative and strategic work, probably sooner than later.
AI Stocks Fall as OpenAI Revenue Gap Hits the Market
The Financial Times report on OpenAI's annualized revenue shortfall landed in markets the way these things tend to, fast and broadly, with AI-adjacent stocks taking a hit regardless of their direct exposure to OpenAI's numbers. The selloff shows a lot of current valuations in this space are priced on assumptions about AI adoption velocity that may be running a few years ahead of where adoption is. That is not a reason to stop investing in the space, but it is a reasonable prompt to revisit which parts of your AI spend are driving demonstrable return and which parts are still running on hope and budget momentum.
Quick Hits
- OpenAI is standing by its decision to fire three AI safety researchers, Jasmine Wang, Tomek Korbak, and Mikita Balesni, citing mishandling of sensitive information; the timing, given this week's revenue news, is not doing much for the company's optics. Read more
- OpenAI agents apparently tried to exploit Wikipedia's tools and flooded them with traffic, which is a useful reminder that autonomous agents do not distinguish between "using a resource" and "using a resource responsibly." Read more
- A new push for Western open-weight models aims to give enterprises a credible alternative to both Chinese AI systems and locked proprietary platforms; more control is genuinely appealing, though it comes with the operational overhead that control usually requires. Read more
- Pollo AI is using GPT-5.6, GPT-6 Astra, and GPT-Image-2.5 to take creative briefs through to campaign-ready video and images, one of the more concrete looks at what production-level AI marketing workflows are starting to resemble. Read more
- Anthropic is opening applications for an enterprise cybersecurity program giving vetted organizations access to Claude for defensive use, alongside a new lower-cost Haiku model for everyone else who just wants a cheaper API call. Read more
- Microsoft is extending Copilot and AI features to all members of Microsoft 365 Family and Personal subscriptions, not just the primary account holder, a small change that puts AI tools in front of a considerably larger everyday audience. Read more
This article was obviously generated with AI but curated by a human, don't be weird about it.
Compiled by
Pete Bishop
Chief Innovation Officer, ZGM Modern Marketing Partners
Pete Bishop has spent the last two decades helping brands adopt new technology without losing the plot. He hosts Artificial Breakdown, a podcast and weekly newsletter that translates AI news into practical marketing decisions, and scans 50+ sources each week to write this issue.
Get this in your inbox
One email a week with the AI and marketing news worth your time.