Researchers studied AI layoffs. Here’s their warning.

Every few months, it seems that another tech CEO predicts that AI will do many of our jobs in the future. But if companies eventually replace too many workers with AI, will there be enough people left with a paycheck to continue buying goods and keep the economy going? In this episode, Katty Kay talks with Professor Gerry Tsoukalas, who recently co-wrote a paper with Professor Brett Falk titled “The AI Layoff Trap.” They built a mathematical model to game out just how bad an AI layoff wave could get — and whether anything can actually stop it. According to their model, many companies may see this firing-your-customers trap coming and walk into it anyway. — BBC Global video description

TL;DR

A ~10-minute New Normal with Katty Kay episode (BBC Global, 20 Jul 2026) in which host Katty Kay interviews Professor Gerry Tsoukalas, co-author (with Professor Brett Falk) of a ~60-page mathematical-economics paper titled “The AI Layoff Trap.” The paper models a specific game-theoretic mechanism, not a general employment-effects survey:

1. The trap, in a monopoly. One firm, one CEO: adopting AI would let the firm cut costs by laying off ~90% of staff, but those same staff are the firm’s customers. A single monopolist internalizes this cost and self-corrects — automates just enough, not more. In isolation, the incentives are aligned.
2. The trap, in competition. With many firms and many workers, each individual CEO’s calculus changes: “my workers weren’t only buying from me anyway — they spread their spending across a hundred competitors.” So the perceived cost of laying off my own workers approaches zero, even though in aggregate, every firm laying off workers the same way erodes the entire market’s purchasing power. This is a classic dominating (dominant) strategy / prisoner’s-dilemma structure: automating is individually rational for every firm regardless of what competitors do, yet collectively it can hollow out demand for everyone, including the automating firms.
3. Why CEOs can’t simply agree not to do it. Even a room full of CEOs (Tsoukalas names Anthropic, OpenAI, Microsoft, Google as an illustrative example, not a reported event) who verbally agree to slow down can’t hold the line once they leave the room — the individual incentive to defect and capture first-mover gains reasserts itself immediately. Voluntary CEO self-restraint is predicted, not observed, to fail — Tsoukalas frames this via the Odysseus-and-the-sirens myth: you don’t trust yourself to resist in the moment, so you need to be bound in advance by an external mechanism.
4. The proposed fix: a tax on full replacement, not on augmentation. Having considered roughly six popular proposals (including universal basic income and worker equity stakes), Tsoukalas and Falk conclude only one directly addresses the incentive itself: make it costly to fully replace a worker with AI, while leaving AI-augmentation cost-neutral — structurally analogous to a carbon tax (doesn’t ban pollution, prices the externality). Explicitly floated alternative framing, given “tax” is politically toxic: subsidies for firms that retain workers rather than a penalty for firms that don’t.
5. The stakes, and the uncertainty. Tsoukalas’s own prediction if nothing is done: “a tremendous amount of wealth inequality… something we’ve never seen before,” potentially leading to political instability comparable to the 2008 financial crisis’s aftermath — but he explicitly flags this as contested (“no one agrees about this question; half the people are on one side, half on the other”) and his own view, not the paper’s proven conclusion.

What was actually ingested

Full ASR-cleaned transcript (~10:17 runtime, 9 named chapters, ~80 dialogue segments), auto-generated captions used (a manual en-GB track also exists per the raw file but was not the fetched track). Single guest, single host, no B-roll content beyond the interview itself.

Linked entities and concepts

• Channel/publisher: BBC — promoted on this ingest (second-source rule: a prior BBC source, the June “AI washing” episode, already existed with no entity page; this second appearance crosses the threshold).
• Promoted Gerry Tsoukalas and Brett Hemenway Falk (2026-07-21, second-source rule) once the underlying paper was ingested — both were dangling, single-source mentions here.
• Dangling (single-source mentions, deferred): Katty Kay (host — also hosted the June episode, so this is her second appearance across two source pages; promote on a third BBC/New Normal source).
• Concepts: ai-employment-effects (primary), dynamic-capabilities (not tagged — see below), systems-thinking (the feedback-loop/collective-action structure of the trap is a systems-thinking-adjacent framing, mentioned in body prose only, no frontmatter tag added since the source itself doesn’t use systems-thinking vocabulary).
• No dynamic_capabilities: tag applied. This source is a macro-economic/game-theoretic model and policy debate, not an account of a specific firm’s digital-transformation process — forcing a W&W cell here would stretch the vocabulary past what it’s meant to classify (per CLAUDE.md’s “when not to tag” guidance).

Relationships

See frontmatter. Two typed supports edges: one to the June BBC “AI washing” episode (same series/host, same broad topic, differently-scoped question — held as a shared-topic pair rather than a contradiction); one to BCG’s reshape-not-replace thesis (descriptive claim about current firm behavior vs. this source’s predictive game-theoretic model of where competitive pressure could push firm behavior). One typed depends-on edge, added 2026-07-21: the underlying paper — this video is a popular-press summary of it, and the paper carries substantial qualification (a reversal case, a post-labor limit, and a six-instrument policy comparison table) the video omits entirely. Considered but not linked: Brynjolfsson Canaries (shares the general over-automation concern but no specific shared claim, mechanism, or data point — too thin for a typed edge; the paper itself does cite Brynjolfsson directly, see that source’s own relationships).