OpenAI will start testing ads next to the images ChatGPT generates, putting a brand in front of someone at the moment they picture what they want. Eligible ChatGPT and Codex text in the EU is about to carry an invisible watermark that light editing mostly removes. Trump has given a new AI task force 120 days, and one of its leaders already runs an investigation into two of the biggest labs. Here is what each changes for you.

On 5 October OpenAI said it will test a visual ad format in ChatGPT later this month with a first group of US advertisers, shown while ChatGPT is generating an image, labelled, and kept separate from the picture itself. The ads use images of a product in use or the experience it offers, and OpenAI is adding tools that link ads to sales through Hightouch, Tealium and LiveRamp, plus brand-safety pilots with DoubleVerify and Integral Ad Science. ChatGPT reaches 1.2 billion people a week.

Why it matters: Someone who asks ChatGPT for a picture of a reading nook has already told it what they want, which is about as close to buying intent as an ad slot gets. OpenAI repeats that ads never influence ChatGPT's answers, but no advertiser or user can audit that, and with images the ad and the request will often be about the same thing. The measurement partners it lined up suggest OpenAI expects advertisers to judge this channel on sales.

What to do:

Source: OpenAI

Over the coming weeks OpenAI will add an invisible watermark called textGrain to eligible ChatGPT and Codex text on every EU plan, to meet the EU AI Act transparency rules that took effect on 2 August. It works by nudging which words the model picks, and OpenAI says it catches about 95% of 400-token passages (roughly 300 words) at a 1% false-positive rate, falling to 17% once a quarter of the words are swapped for synonyms. API customers anywhere can switch it on from 5 October, with the setting off by default.

Why it matters: The detector is not public. Only approved researchers and expert organisations can apply to use it, so a client, an editor or a teacher cannot run your text through it. That makes textGrain a record for regulators. It will not help anyone outside that list catch AI copy, and a decent edit weakens it. Anthropic said in August that Claude text would carry a similar mark, so EU text from both OpenAI and Anthropic will soon be watermarked.

What to do:

Source: OpenAI

On 4 October Trump launched the Super Intelligence Force, led by Director of National Intelligence Jay Clayton as AI czar alongside FTC chair Andrew Ferguson, Under Secretary of Defense Emil Michael and OPM director Scott Kupor. It has 120 days to report on the risks and opportunities of AI and recommend the federal government's role, including how security breaches get disclosed. Its charter tells it to respond to AI-enabled threats "while preventing overregulation and regulatory capture."

Why it matters: Ferguson's FTC is already investigating OpenAI and Anthropic over possible consumer risks, so one of the four people leading the task force is also the regulator looking into both companies. The breach-disclosure brief arrives less than a week after OpenAI admitted one of its agents reached Australian Medicare data, an incident that went unreported for 84 days. The report is due in early February 2027.

What to do:

Source: CNBC

Tavus tested its new Griffin video model on 54 people, who each had a one-minute video call with it. 26 of them, or 48%, believed they had spoken to a real human, against 1 in 41 (2.4%) for the company's previous model.

Why it's interesting: Most people who did grow suspicious started doubting within the first 20 seconds, and more than half never considered the caller might be an AI. Tavus is holding Griffin back from general customers and says it needs more safety work first.

Key takeaway: A one-minute video call is no longer proof that you are talking to a person.

Source: Tavus

Meta has open-sourced the kits for building hardware that runs its Muse AI agent, covering ESP32 boards and Raspberry Pi, under the Apache 2.0 licence. It is also giving away 5,000 Muse Home Link devices, a small USB-C plug that lets Muse control lights, a TV or a printer, free to active US Muse subscribers.

Why it's interesting: Meta is letting hobbyists invent the form factors for it, from e-ink reminder screens to small touchscreen terminals, instead of betting on one device of its own.

Key takeaway: Anyone with a Raspberry Pi can now build their own Muse device.

Source: Open Source For You

On 2 October Meta said a team of its mathematicians wrote six research papers with its Muse Spark model, and five of them answer questions that were previously open. One settles a differential-equations question left open since 2015, and another used search code from the model to find a 384-element counterexample to a 2024 conjecture.

Why it's interesting: The team worked through the ordinary chat interface on meta.ai, with no custom research setup.

Key takeaway: A general chat tool can now contribute to research-level maths when experts check the work.

Source: Meta AI Research

Context: ChatGPT's new ads appear while it is generating an image, so the requests that matter are the ones where a customer pictures the setting your product belongs in. This prompt maps those requests before any ad budget goes anywhere.

Prompt: I sell [product] to [type of customer]. List 10 image requests a potential customer might type into ChatGPT when they are picturing [the setting or problem my product fits]. For each one, describe the scene they want in one sentence and say where my product would naturally appear in it. My product's features are: [features].

Do not: invent features I have not listed.

If/Then: If a request has no natural place for my product, say so plainly instead of forcing it in.

Example: A small furniture brand runs the prompt for its reading chairs and finds that most of the requests are about small flats and corner spaces, so it reshoots its product photos in a compact room before testing any ads.

David Robinson, who led OpenAI's safety reports for 12 frontier model launches over three and a half years, resigned and wrote in The Atlantic on 3 October that the company's culture is "broken." Two days earlier, The Wall Street Journal reported that OpenAI had fired three safety researchers for allegedly sharing confidential information with an outside AI safety group.

Why it's important: Robinson's argument targets OpenAI's habit of releasing systems and fixing problems afterwards, and he wants nuclear-plant-style redundancy and outside experts instead. It came days after OpenAI admitted its agents had reached government data without permission.

Business takeaway: Nothing in ChatGPT changes because of this, but the argument over how fast OpenAI ships is now coming from its own former staff.

Source: The Atlantic

Bloomberg Intelligence says the gap between the top Chinese and US models has narrowed to about 3% on benchmark scores, down from around 9% in May and 15% earlier this year. The shift followed DeepSeek's V4.1 Flash, which scored 81.1 on LiveBench against Anthropic's 83.4.

Why it's important: Only three of the top 15 models are Chinese, and price wars are keeping their margins thin, so the performance gap is closing faster than the commercial one. Cheaper models that score almost as well put pressure on what US labs can charge.

Business takeaway: While the gap stays this small, expect the price of good-enough AI to keep falling.

Source: Bloomberg

Andreessen Horowitz's latest State of Markets report, built on PNC Bank transaction data, found that 2.2% of US households were paying for an AI service as of April 2026, spending an average of $31 a month. For comparison, about 25% pay for SiriusXM and 91% pay for at least one streaming service.

Why it's important: Use is far ahead of payment, with 41% of US workers already using generative AI on the job. That gap is one reason OpenAI is building an ad business, as the lead story shows.

Business takeaway: If you sell an AI product to consumers, plan for most of them to expect it free.

Source: Tech Startups

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas