Meta is sued over AI layoff scoring, and entry-level hiring keeps shrinking

A performance score built by scanning keystrokes, screen activity, emails, and browser history has one blind spot: it cannot record work that was never done because the worker was legally absent. That is the heart of a lawsuit filed late Monday in federal court in Oakland by 26 former Meta employees.

They were among the roughly 8,000 people, about 10% of Meta's staff, told in May their jobs would be eliminated. Each of the 26 had taken protected medical, parental, or family leave, or asked for a disability accommodation. The complaint says Meta ranked workers for the cuts using internal AI systems, activity-monitoring data, token-usage dashboards, and algorithmically assisted performance rankings, and that such scores, by design, cannot be accumulated by someone on leave.

A system that measures output over several months, then ranks people by it, reads a parental leave as a run of low output; it cannot tell the parent who was legally away from someone who did little while at their desk. Because women take most pregnancy and caregiving leave, the plaintiffs argue, the ranking fell harder on them.

Meta rejects the claim, saying workforce decisions "were and are made by people, not AI." The workers are asking the court to pause their separations, set for July 22, while they pursue arbitration. They also say Meta never tested its scoring systems for bias, which they argue breaks recently adopted laws in California and New York City that require exactly that check.

The share of large companies planning to cut entry-level roles jumped from 17% to 43% in a single year, a survey of 415 chief executives by the Oliver Wyman Forum and the New York Stock Exchange found. The reason they give is automation.

Junior work is the easiest to hand to a model: research, scheduling, data pulls, and first-draft writing, the routine tasks a new hire used to learn on. When those go to software, the roles built around them stop being posted. In the United States, employment for workers aged 22 to 25 in the jobs most exposed to AI has fallen about 16% since late 2022, by a June report from the World Economic Forum and PwC.

The immediate worry is youth unemployment. The longer one is the pipeline. Every senior analyst or manager a business promotes in five years is someone it hires and trains now, and cutting the bottom rung thins the supply of experienced staff later, across the whole market.

Not everyone is cutting. IBM plans to triple its US entry-level hiring this year, and among the companies getting the strongest returns from AI, 24% say they are hiring more junior workers, not fewer.

What does it cost to run an AI agent? The price of a token is the wrong answer, OpenAI told company leaders in a guide published July 14. Headline prices have fallen fast: the company says the price per million tokens dropped 97% from GPT-4 to GPT-5.4, and its newest model, GPT-5.6, does better on a coding benchmark while using 54% fewer output tokens and 57% less time per task.

But a cheaper model that fails, retries, or needs a person to correct it can cost more in total than a stronger one that finishes in fewer steps. The measure OpenAI recommends is cost per accepted outcome: in customer support, a case actually resolved; in engineering, a code change that passes review.

From September 15, Cloudflare will change a default that most website owners never set. On sites it protects, AI crawlers that gather data to train models, and the agents that fetch pages in real time for a user, will be blocked automatically on any page that carries advertising. Ordinary search crawlers stay allowed.

Cloudflare handles traffic for roughly a fifth of the web, so the switch reaches far. It splits AI visitors into three kinds: search bots that index a page to answer questions later, agents acting live on a person's behalf, and training crawlers that absorb content into a model. On ad-supported pages, the last two lose access by default.

For a business whose research, monitoring, or customer-service tools pull live information from the open web, the practical effect is a narrower set of pages those tools can read. The change applies to new sites and to existing customers on the free plan; anyone can opt out in their settings before September 15.

Apple released a free public beta of iOS 27 on July 13, opening its rebuilt Siri to anyone with a recent iPhone. The assistant now reads a user's mail, messages, and photos, responds to whatever is on the screen, acts across apps, and answers general questions the way ChatGPT or Gemini does.

It is Siri's largest change since 2011, and it needs an iPhone 15 Pro or newer to run. It also arrives in English only, and Apple is not offering it on iPhones in the European Union at launch, with no date given. The full release is expected in September.

For a team whose staff use iPhones, that splits the tool by geography: the same phone will do materially different work depending on the country its owner is in.