Keep your own answer in the room

AI can influence a decision before anyone clicks “approve.” It can also influence the story we tell about an entire labor market, or the gains a company expects from smaller teams. This issue is about keeping enough independent judgment to notice what the recommendation, or the headline, leaves out. A human signature doesn’t tell us much about how much thinking happened before it.

Human review can inherit the machine’s bias

A University of Washington study found that 528 participants’ candidate selections followed racial biases in simulated AI recommendations.

Our take: “A human makes the final decision” is an incomplete answer to a very real problem. If the machine supplies the shortlist and the person mostly agrees, you may have added a signature to its bias. We would want reviewers to form a view from the evidence before seeing the recommendation, with enough authority to reject it. That design still needs evaluation; a person in the workflow is no automatic guarantee.

AI exposure and AI adoption tell different stories

Revelio Labs’ labor-market analysis distinguishes jobs whose tasks AI could perform from firms classified as adopting AI through their hiring.

Our take: Be suspicious of any chart pressed into proving that AI either destroys jobs or creates them. An expanding company can hire more people while opportunities for junior workers narrow. Revelio also notes that adopters were already growing faster, which should take some swagger out of the causal claims. Workforce planning needs that messy detail more than it needs a winning side in the argument.

More code doesn’t establish a case for smaller teams

Reuters’ investigation of Meta’s AI restructuring describes the company canceling planning for a second restructuring wave amid employee backlash and internal reports questioning AI productivity gains.

Our take: Meta’s internal figures showed software-platform code changes up 220% year over year, while changes delivering user features rose 36%. Those are different activities, so dividing one by the other wouldn’t give us a productivity measure. They do raise a useful question: where did the extra activity go? Before redesigning a team around assumed AI gains, look at what reaches users and the work required to keep it running. Reuters couldn’t establish exactly why Zuckerberg changed course. We would want staffing assumptions to rest on demonstrated gains, with employees helping identify what the activity counts miss.

One useful move

Have two possible approaches to a project? Write down which you currently favor and why, then give both options to your usual AI tool:

Option A: {{option_a_details}} Option B: {{option_b_details}} Additional context: {{decision_context}} Make the strongest practical case for each option. Identify the assumption each depends on most. Use the details I’ve provided, and flag missing facts rather than inventing them.

Compare the answer with your original view. Add any useful objection to the project note you’re already writing.

Something fun

Ask your chatbot to convene an expert panel about a genuinely trivial decision:

A pirate, a museum curator, and a very tired golden retriever must choose a name for a neighborhood sandwich shop. Give each a different favorite and one wonderfully specific objection to the others. End with the compromise name nobody loves.

The golden retriever should have excellent instincts and absolutely no interest in the branding deck.

Built in the past two weeks

We published a getting-started guide for the Mastermind’s Open WebUI workspace. It walks members through finding custom assistants, comparing models, using saved prompts, and setting up personalization and memory. The aim is to give someone a useful first route through the workspace before they start customizing everything.