26 of 54 people thought the AI on the video call was real

Issue 007 · October 1 to October 7, 2026

In this issue

Tavus released an AI that holds a live video call, and 26 of the 54 people who took one thought it was a person. OpenAI began watermarking ChatGPT's writing, then published the trick that erases it. And a16z found almost a third of big companies claim AI results. About 2% track a number. Plus: ask your AI what it doesn't know about you.

Dave's Read

On October 1 Tavus released a video model where a video avatar carries on a live face-to-face call. In the study it published, 26 of 54 people tested came off a one-minute call believing they had talked to an actual person.

That 48% should scare us all. Business owners and employees already treat a face on a screen as proof of who somebody is. But what if a vendor calls to say their bank details changed? Or somebody who looks like your bookkeeper asks for a login. The fact that you recognize the person on the screen no longer means that the face on the screen belongs to them.

Griffin runs the speech, the voice and the face animation at the same time instead of one after another, which is why it can interrupt and be interrupted.
One of the loudest replies said this kind of thing should be illegal.
Emad Mostaque's read was shorter: remote work is cooked.
You can't buy it yet. Griffin-Lite is a research preview for selected testers, and nobody has run the test independently.

What this means

A face on a video call is not proof of who is on it.

None of this means you need new software. The thing that fails here is a habit, and the fix is a rule you write once. The FBI has printed the same sentence about wire fraud since 2022: use a second channel to check any request to change account information.

The news worth your time

Tools

OpenAI will hide a watermark in ChatGPT's writing, and a light rewrite takes it out

OpenAI · October 5

Announced October 5, for ChatGPT and Codex users in the European Union, on every plan. The reason is Article 50 of the EU AI Act, which says AI-written content has to be marked so software can spot it.
OpenAI published where it breaks, which is the useful part. Swapping 10% of the words for synonyms dropped detection from about 92% to 66%. Swapping 25% dropped it to 17%.
The detector stays private, going to approved researchers case by case. So if somebody accuses your proposal of being machine-written, no tool settles it (and the one that might isn't a tool you can get). It beat other methods, including Google's SynthID.

Measurement

Almost a third of big companies claim AI results. About 2% can show a number they track

a16z · September 30

Andreessen Horowitz published State of Markets II on September 30. From page 27: "While nearly 30% of SP500 companies report some 'quantifiable impact' of AI, only ~2% are reporting any tracked metric."
To be fair, a tracked metric and a press release are different jobs, and plenty of these companies measure things they never publish.
Even so, these are businesses with investor relations departments and a legal duty to disclose. Reporting numbers is the work.

Of S&P 500 companies, the share reporting an AI metric they track over time. Nearly 30% claim quantifiable impact. a16z, 2026.

If you're not measuring what AI changed, how do you know it's working?

Skim these, then forget them

Shopify put store design on one canvas you talk to. In early access. It matters if you sell online and not at all if you don't.

Amazon pledged more than $1 billion over five years to the towns its data centers sit in, for community college, trades training and energy bills. A real number that reaches your business through the news and nowhere else.

Worth following

One for the drive. The other is for when you've had enough of AI.

The AI in Business Podcast is Emerj's weekly interview show for leaders who don't write code, hosted by Daniel Faggella. Fair warning: the guests come from very large companies, so take how they think and leave their budgets.

Simon Sinek is here for leadership, not AI. Put it on when the week has been all agents and watermarks.

For the further along

Skip this unless money or logins change hands on somebody's word.

Write one sentence naming what gets confirmed on a second channel: bank detail changes, wire requests, password handovers.
Pick the channel. A number you already had, never one from the message asking.
Tell the two or three people who can send money. That's the rollout.
While you're there, open your Google Account's linked apps and pull anything you don't recognize.
If it already happened, report it to the FBI's IC3.

Try this week

Fifteen minutes. Do this one.

Ask your AI what it doesn't know about you.

Type this: "What would you need to know about me and my business to be more useful?" Then answer it.

Read the list before you answer it. That's what you've been leaving out of every prompt since you started. Then save your answers where the tool can see them.

The prompt

This one teaches the bad example trick. Showing beats describing, and one example outruns a paragraph of instruction.

Here is an example of exactly what I do not want: [paste it]. Now write the opposite, for this: [your thing].

Reach for it when you can't explain what you want. You'd know it when you saw it.

One reply

Hit reply and tell me one thing. What did it ask for that you hadn't thought to tell it?

The Kitala Brief