Skip to content
Business Efficiency Talking point

TypeSafe AI's Almeida argues unreliable AI pushes checking costs onto other people

Diogo Almeida argued that firms avoid trusting AI with costly decisions, so others absorb the cleanup.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
TypeSafe AI's Almeida argues unreliable AI pushes checking costs onto other people

Photo: https://kaboompics.com/ / Pexels

In brief
  • Diogo Almeida argued that AI is unreliable, so firms keep it away from decisions that would cost them and push that cost onto others.
  • He said slop is what results when nobody reads and filters the output, and bad actors gain most from that.

Diogo Almeida, cofounder of TypeSafe AI, argued that AI "slop" is a problem of who pays. Because model output is unreliable, he said, firms cannot use it for decisions that would cost them. So the checking and cleanup land on other people. He made the case in a talk at CODE Summit 2026, featured on the AI Engineer show.

What was said

In Almeida's account, slop starts with unreliability. A company using AI cannot lean on it for decisions where a wrong answer costs the company.

He described the way out bluntly: "it's actually totally fine to push the costs into other people, which is horrible, externalized costs." An externalized cost is one that someone other than the maker ends up paying.

Slop, in his telling, is what that looks like in practice. If people read the output, filter it and care about quality, it is no longer slop. The world is full of it, he said, because bad actors who do not care hold an asymmetric advantage.

He also described the failure mode. Of current models he said: "you get Einstein 95% of the time, but Mr. Bean 5% of the time." Nothing in his account marks which answers are the weak ones. His question was how anyone can automate anything if you cannot tell which is which. The implication is that every answer needs a check.

His alternative is narrow, consistent reliability, so the model is right every time, rather than occasional brilliance that a person must verify. He said that kind of reliability does not show up on benchmarks.

Why it matters

Our reading: this reframes slop from a taste problem into a budgeting problem. A firm that cannot afford a wrong answer on its own decisions may still send out unchecked text, code or messages when the error costs fall on the recipient.

For managers and engineering leaders, that raises two questions. Where does unchecked AI output leave your team and become someone else's job to verify? And where does the same flow arrive at your desk, from vendors, applicants or contributors?

The other side

Almeida offered a diagnosis and a direction, not evidence of how large the shifted costs are. In the excerpts we have, he gave no figures and named no firms or victims.

His answer is also tied to his own company's bet on narrow, reliable models, so readers should weigh it as one vendor's view.

He also allowed that output people read and filter is no longer slop. The open question is who bears the checking cost when the sender does not care.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from