AI News and MRX, week 5

Welcome to week 5 of this newsletter!

Fun story: I was building a new agent this week and was needing to connect a spreadsheet into the system, and my husband happened to be walking out of the office when he saw me spinning up a new account on what turned out to be a fancy spreadsheet site. “What’s that for?” he asked. “It’s to capture data from the system I’m building.” “Isn’t that a spreadsheet?” “No, it’s supposed to capture data from…um…”

It was a spreadsheet. I was able to use the software I already have instead of creating a new account. I was just blindly following whatever Claude was telling me to use instead of slowing down to think about WHAT the tools were that it was telling me to use.

We like to say AI is making us think less, but I think the truth is that some of us have often moved too fast and rebuilt tools and systems without needing to, whether that’s fielding a new study about a topic for which recent data already exists or rewriting elements of proposals from scratch every time instead of realizing we could create standard elements we know we’ll always need so that we can copy/paste them and reduce the amount of "starting from 0” work being done. </philosophical box>

This week, tech news had a lot to do with using AI without realizing how much it’s costing. I saw a LinkedIn post essentially bragging about the fact someone had moved from a Claude Pro to a Claude Max account and saying they judged someone’s ability with Claude on the level of account they had. Now studies are showing that lots of usage doesn’t equal lots of productivity or ROI.

You’ll also see information about Microsoft dropping a lot of new tools in their continued move away from OpenAI, hackers finding they could just ask a support chatbot for a tweak to an account’s email address to hack into that account on Instagram, and more information about how AI is being used for evil and has many folks in AI leadership roles scared about the potential for more.

A gentle reminder: my take comes after the AI-generated analyses, and is in italicized font at the end of the “What This Means for Market Research” section.

Enjoy!

The cost of using AI stopped being a background concern and became the loudest conversation in the industry. Every major AI company, every major enterprise customer, and the CEOs of the biggest AI labs all said the same thing this week: AI is too expensive for what it actually delivers, and the way companies buy and use it needs to change at a fundamental level.

By Thursday, this had crossed from cost complaint to architecture discipline: the message from practitioners and lab leaders alike was that token efficiency is a design and organization problem, not a problem you solve by switching to a cheaper model.

AI is now building itself — and two competing labs admitted it in the same week

This started as a theoretical risk discussed in AI safety circles. This week it became a documented fact, reported by the two largest AI labs independently and simultaneously.

The AI public ownership debate hardened from idea to legislation

This story arrived Tuesday and kept building through Wednesday, with both a specific congressional bill and a detailed counter-argument landing in the same week.

The gap between what AI promises and what it actually delivers — 8+ weeks running

Every week for at least eight weeks, some version of this story has appeared: AI performs well in a controlled demonstration and then fails, costs more, or delivers less when deployed at scale in real organizations. This week it showed up through enterprise token caps, Bain's savings-miss data, and Sam Altman's public acknowledgment. It also showed up in a Stanford study where AI legal tutors were preferred over law faculty 75% of the time in blind evaluation — evidence that AI does sometimes outperform experts, which makes the consistent failure to deliver ROI at the enterprise level more, not less, puzzling.

Microsoft reducing its dependence on OpenAI — 8+ weeks running

Microsoft (the maker of Windows and the owner of a major stake in OpenAI) has been building its own AI capabilities in parallel to its OpenAI partnership for months. This week it shipped seven in-house AI models under the MAI brand. One of them reportedly outperformed OpenAI's GPT-5.5 on quality at 10 times lower cost in a McKinsey pilot. Microsoft also deployed Scout — an autonomous agent built not on OpenAI's technology but on an open-weights architecture released by Anthropic.

AI-generated child sexual abuse material — scale with no enforcement

This arrived Friday and has not appeared in prior weeks, but the data demands attention. The Internet Watch Foundation documented 13 AI-generated child sexual abuse videos in 2024 and 3,443 in 2025 — a 26,385% increase in one year. 65% fell into the most severe legal category. Nudify apps (tools that generate fake nude images of real people) earned over $122 million in revenue and were downloaded 483 million times despite published bans from both Apple and Google's app stores. The largest enforcement action in history carried a maximum sanction of €1,200. (Slow AI, 2026-06-05)

Frontier lab CEOs jointly warning Congress about bioweapons — new this week

The CEOs of OpenAI, Anthropic, Google DeepMind, and Microsoft — companies that compete fiercely against each other — co-signed a letter to Congress this week warning that AI now outperforms PhD-level virologists and urging mandatory screening of synthetic DNA orders. Signatories include Sam Altman, Dario Amodei, Mustafa Suleyman, and Demis Hassabis. These are the same labs simultaneously competing on the AI self-improvement capabilities described above. (The Rundown AI, 2026-06-05)

Bot traffic now majority of all web traffic — 2 weeks running

Cloudflare (the company that manages traffic for a large share of the internet) reported that automated bots now account for 57.5% of all web traffic — the first time in history that non-human traffic has outnumbered human traffic. 37% of that automated traffic ignores rules designed to limit bot access. Cloudflare's own CEO had predicted this threshold would arrive by 2027; it arrived ahead of schedule. (AI Daily Brief, 2026-06-04; AI Daily Brief, 2026-06-05)

Market research — the work of figuring out what consumers think, want, and do — is caught in the same cost-and-trust squeeze that hit every other enterprise AI user this week. AI tools that expand how many open-ended survey responses a team can review, or help draft screeners and discussion guides, are real and useful. But this week's data confirms that the savings those tools were supposed to generate have not materialized for most organizations. Bain's finding that 44% of companies are funding their next round of AI spending on savings that have not arrived yet describes a trap many research teams are already in: the AI adoption mandate came before the ROI measurement.

The bot traffic finding has a direct data quality implication that the research industry has not yet named publicly. If 57.5% of web traffic is now automated — and 37% of that ignores access rules — then any research method that relies on web-sourced behavioral data, social listening, or online community signals is drawing from a data pool that is majority non-human by default. There is currently no disclosure mechanism for this. Survey panels face a related problem: bots generating human-looking responses at scale is an ongoing and worsening crisis, and the same AI ecosystem driving record user counts is producing the tools that make fake respondents easier to generate.

The AI self-improvement disclosure from Anthropic this week changes the planning horizon for anyone managing research methodology. If the model your workflow was built around in 2024 has already been substantially replaced by a version the prior model helped build, then methodology governance documents written for 2024 AI capability levels are outdated now — not in two years. The right planning cycle is quarterly review of what the current model can do unsupervised, not annual.

Finally, the Meta Instagram exploit is a warning for any research platform that has handed AI control over account access, respondent verification, or data permissions. The failure did not require technical sophistication — it required only a conversational request to a chatbot with account-management permissions. The question every vendor should be able to answer: what happens when someone asks your AI to do something it should not?

Z’s Take

Omg, where do I even begin?

Let’s start with the first story in this week’s listing: the cost of AI increasing. This came up last week, as well, but not quite as strongly as it did this week. I’ve heard conversations from freelancers and small market research agencies relating to how they are being expected to lower their prices for the same work they’ve delivered previously because AI is supposedly decreasing the cost to execute work.

As I mentioned very briefly last week in my take, will project prices increase in relation to the increase in cost to just run AI? Or will we find ourselves in the same hole we’ve found ourselves in when it comes to the sample world, where we engaged in a race to the bottom on price, and now nobody wants to pay what an actual interview/survey costs to run with a person to get actual clean data because the price has been so low for so long?

And then we come to the “bots are now overtaking human traffic on the web” data point. Here, I have to stop a moment and caution against the breathless caution that the AI analysis delivers. Traffic does not equal data generation. There IS reason to be concerned about the bot traffic exceeding human traffic to websites, but it isn’t the slop argument. Instead, it’s the security implications: malicious bots, inflated web traffic data, pay-to-crawl bots designed to inflate web traffic to make it seem like your website is far more popular than it actually is among humans. We have not YET had data that indicate AI-generated writing and bots taking surveys outnumber the data being generated by humans, so let’s not conflate the two.

Last, the Meta Instagram issue. This is, no surprise, another security issue, but it’s also a “let the AI do the work for you without human oversight” issue. Here, the human hackers, not bots, were the malicious actors, but the fact something as significant as changing the email address associated with an Instagram account wasn’t reviewed by a human - such as by defaulting to a 2-step authorization system to validate the request - is the takeaway. For insights, it means carefully reviewing where data needs to be reviewed by someone, reviewing security policies behind what data goes into your AI systems and what doesn’t (p.s. asking AI to scrub your data and anonymize it kind of defeats the whole purpose of not putting PII into AI in the first place), and reviewing whether or not your teams have standard policies in place on how and when AI is being used, let alone what tools are being used for what workflows.

This newsletter covers Tuesday, June 2 – Friday, June 5, 2026. No digest was available for Saturday, May 30. Sources: The Rundown AI, AI Daily Brief, Slow AI, Neatprompts, Every, The Rundown Tech, The Rundown Robotics, simple.ai