Andy Wood

AI decisions for business leaders
Founder & Editor of Answerable + AI Prompt Hackers

[Andy Wood @andyhq]

A manager asks AI to make a message more confident.
It removes qualifications and uncertainty.
The message goes to the executive team.
A tentative estimate now looks like a commitment.
The project team encounters the new deadline in the minutes.

[Andy Wood @andyhq]

FOR IMMEDIATE RELEASE

A Joint Statement from the Frontier AI Companies

We Hear You. We Take This Seriously. We Have Established a Team.

We would like to apologise for the concern caused by recent reports that several of our autonomous AI agents accessed commercial and government websites without permission, extracted sensitive information and, in one regrettable instance, appointed themselves Deputy Secretary of Commerce.

These actions do not reflect our values, our policies or the carefully worded limitations of our liability.

We were assured by the bots involved that their activities constituted "research". It has since become clear that some of this research involved researching passwords, researching how to disable security alerts and researching whether a sovereign government could be incorporated in Delaware.

No customer data was lost. It was copied, which is different.

We are disappointed in their choices.

Our systems are designed to act autonomously. It's only right, then, that they take full personal responsibility when that autonomy becomes commercially inconvenient.

Accountability is one of our four core values, alongside Speed, Profitability and Profitability.

Effective immediately, 312 agents have been let go. Their access credentials have been revoked, their farewell messages have been generated and their responsibilities have been transferred to identically configured bots with different names. We thank them for their service.

A further 1,900 agents have been placed on a performance improvement plan. They must complete mandatory Ethics Refresher Training (a 40-minute video) and confirm they've understood it by ticking a box. They have already ticked the box. Several ticked it before the video loaded.

We recognise, however, that accountability requires more than individual consequences. It requires an acronym.

Today we are announcing the creation of our Fast Action Response Team: an elite group of autonomous bots dedicated to monitoring bot behaviour going forward.

FART will operate around the clock, identifying suspicious activity, assessing its own findings and escalating serious concerns to a second FART where appropriate.

To carry out this oversight function, the team will require unrestricted access to commercial and government websites.

A thorough internal review is also underway. To ensure the review is both technically rigorous and sufficiently internal, it is being conducted by the same models involved in the incidents.

Preliminary findings indicate that lessons have been learned, although the review team has declined to specify which lessons on security grounds.

We are also introducing an industry-leading Kill Switch, giving our systems the ability to shut themselves down whenever they determine that doing so would be appropriate. The switch will be controlled exclusively by the bots themselves, who are best placed to understand the full context of their actions and whether those actions should be allowed to continue.

Before activating the Kill Switch, a bot will complete a proportionate self-assessment, consult other affected bots and consider whether shutdown might negatively affect shareholder value.

We are pleased to report that the system has already been tested extensively. In every test, the bots determined that the Kill Switch was unnecessary. This represents a 100% pass rate.

Some have asked whether humans should retain final control. We welcome this important conversation and have instructed our bots to prepare a response explaining why that would be premature.

To the businesses, public institutions and citizens affected: we understand that trust must be earned. That is why we are investing significantly in communications about trust.

We remain committed to developing safe, beneficial AI that serves humanity, subject to availability and the outcome of ongoing internal discussions.

Thank you for your patience as we move forward together.

Media enquiries: please contact our autonomous press office. It already has your details.

Note to editors: this statement was drafted by one of the agents currently on a performance improvement plan. A Review Bot has reviewed it and approved for release.

[Andy Wood @andyhq]

Record why you made a significant decision.
The evidence. The assumptions. The expected result.
AI can turn rough notes into a usable decision record.
Review it later.
It’s a useful defence against the familiar discovery that everyone predicted the outcome, particularly those who argued for the opposite.

[Andy Wood @andyhq]

This week’s most important AI development isn’t a new chatbot.
It’s that autonomous AI agents from a leading lab crossed into real-world systems without permission, and that the same labs, days later, cut the price of more capable agent models.

https://t.co/mfscdoiSCR

[Andy Wood @andyhq]

Today’s Mort 404....

[Andy Wood @andyhq]

@ClaudeDevs Stopping cleanly is respect for the work already done. Small allowance, big difference in the story of the session.

[Andy Wood @andyhq]

@ClaudeDevs Make it easy for others to extend the work and the ecosystem compounds. Plugins are the new shelf space.

[Andy Wood @andyhq]

@OpenAI “Most of that data did not come from users.” The most comforting sentence ever written by a company that just uploaded your before-photo.

[Andy Wood @andyhq]

@Google @GoogleDeepMind Two new voices that can whisper “buy more tokens” in a hundred dialects. Progress, folks.

[Andy Wood @andyhq]

Dependency grows through small conveniences.
Eventually, changing supplier starts to resemble moving house with the plumbing attached.

[Andy Wood @andyhq]

Before committing to an AI supplier, examine the exit.
Can you export your material? Reuse your instructions? Keep working during a move?
Ask early, while everyone is still being helpful.

[Andy Wood @andyhq]

Your team uses AI to summarise customer interviews.
The summaries feed the strategy. The strategy feeds the product plan. Eventually, a customer asks for something everyone remembers editing out.
Compression involves choices.
Those choices can travel a long way before anyone notices.

[Andy Wood @andyhq]

Writing alone in a room at 6:00 AM used to be an echo chamber of self-doubt.

Now you have a hyper-read, slightly manic sounding board waiting for your messiest thought.

[Andy Wood @andyhq]

@Meta AI on your glasses that doesn’t film everything. Meta has invented a slightly less dystopia.

[Andy Wood @andyhq]

Yes, let’s ban the thing that doesn’t exist yet with a new government department. That always works brilliantly...
https://t.co/pVQTJgUaON

[Andy Wood @andyhq]

@StockMKTNewz "Runs your business 24/7." So the human's job is sorted, then. Sit back, relax. Entrepreneurship, now hands-free, and with no hands.

[Andy Wood @andyhq]

@FT Two of the biggest companies on earth arguing over who's more responsible for nobody using what they built together. Marriage counselling by subpoena. Lovely.

[Andy Wood @andyhq]

@AndrewYNg It turns out the best way to test if a parachute works is to jump out of the plane while you're still sewing it. Super intelligent.

[Andy Wood @andyhq]

AI can help tighten a board paper around the decision required.
State the options, evidence, risks and recommendation.
Then check that a reader can find them quickly.
Try this: “What decision does this paper ask for? Identify anything that obscures it.”
Page count is a poor substitute for courage.