The Mood on AI: Who is watching the agents?

One developer on Bluesky described a partner company that ordered all of its developers to use AI. Now work stops after three weeks of every month, because the AI budget runs out.

The Mood on AI is a weekly read of how people are talking about artificial intelligence, put together by Fíor from a review of public posts on X, LinkedIn and Bluesky. It is a snapshot of the conversation, not a poll.

This week's mood was mixed, and more wary than hostile. Very few people argued that AI is useless. Very few argued it should be trusted on sight either. The common thread was a condition: AI is welcome where people can see what it is doing, check it, and stop it.

Who is watching the agents?

The most common subject was AI agents, the kind that act on your behalf instead of just answering questions. It was a week of launches for them, and the reaction was cautious.

Security professionals compared agents inside a company to an insider threat: something with access to everything and no one checking its work. One developer described finding a prompt injection flaw in his own agent, and being unable to tell where some of its information was coming from. Others drew hard lines. No access to email or passwords without strong authentication. No sending, paying or signing without a person approving first.

Even the optimists set terms. The most upbeat posts about agents came with conditions attached: limited permissions, independent inspection and a full record of what the agent did.

Does it actually pay?

The second thread was money. Alongside the developer whose AI budget runs dry each month, a marketing consultant questioned whether the promised productivity justifies the scale of investment. An engineer said he had heard believable stories of AI making teams faster, from people he trusts, but still had his doubts.

The positive voices measured AI differently. They talked about hours saved and repetitive work removed, not about jobs replaced. Supporters and sceptics were often arguing about different things.

The human part

Work and craft came up again and again. Some posts worried about AI taking entry-level jobs, the ones people use to start a career. Others pushed back on replacing human voices in marketing with generated content.

One designer shared a Fortune piece about a cymbal maker in Massachusetts, where a person still strikes and listens to every cymbal before it can carry the family name. A product designer made the same point about his own field. AI can now draft interfaces and summarise research, but understanding people is still the job.

Privacy and control

A smaller group focused on control. Would the companies behind AI protect people's data, or use it to make decisions on their behalf? Another post shared a report suggesting students did worse after a school brought in AI tools.

The most memorable version was the simplest. One Bluesky user said that in the age of generative AI, they no longer trust any device with few or no buttons. No buttons, no deal.

How the story is told

Several people took aim at the media itself. Two criticised a documentary that dramatised an AI security incident, saying the storytelling made it hard to separate what happened from what was imagined. Another argued that coverage makes AI sound like magic and skips over the ethical questions.

That matters to anyone who reads about AI, which now means almost everyone. Trust in the technology is tied up with trust in how it is reported.

Where the optimism was

The hope this week was practical. Developers welcomed AI that finds security flaws before attackers do, in the same week AI helped uncover serious flaws in PHP. A scientist pointed to AI speeding up the search for new antibiotics, while warning that the same tools raise biological security risks. Others were excited about open models and about shops preparing for AI shopping assistants.

The reading

Almost nobody this week asked whether AI can do the job. They asked whether they could trust it, and whether they could check. The excitement has not gone. It has acquired conditions.

The people most comfortable with AI this week were the ones who could see inside it. That is a reasonable thing to ask of a machine, and of the people writing about it.

Try the browser extension at Fíor.