AI Kept Sending a Laid-Off Worker to Jobs He Couldn't Lift

Marcus kept a notebook beside the laptop. On the left side of each page, he wrote the job the state-funded career tool recommended. On the right, he copied the physical requirement from the posting.

Warehouse lead: 50 pounds.

Maintenance technician: frequent bending and lifting.

Production supervisor: able to cover any station.

His limit was 25 pounds.

Marcus had worked at a factory for 17 years, most recently running equipment that cut and packed small metal parts. He earned $27.10 an hour before the plant cut a production line in March 2024 and eliminated his job. The work included machine setup, measurements and written checks, but job listings often reduced it to a broader label: machine operator.

A back injury had changed what he could do. He could stand, walk and work a full shift. He could lift up to 25 pounds without repeating the motion. His clinician had written the restriction plainly, and his former employer had adjusted his station during his final year.

The career tool knew none of that.

It appeared inside the online system Marcus used for a state-funded retraining program. A chat window asked about his experience, the work he wanted and the minimum pay he would accept. Marcus entered $26 an hour, close to his old wage, then described setting up machines, checking dimensions and recording production results.

Within minutes, the tool produced ranked occupations and local openings. It said his manufacturing experience could transfer to warehouse supervision, industrial maintenance and production leadership. The explanations were fluent and specific enough to feel considered. One noted that machine setup showed troubleshooting ability.

Another linked his inspection work to safety and quality control.

The recommendations were not random. They were also unusable.

What the ranking could see

The system did more than search for words on a résumé. It converted Marcus's work history and chat responses into a numerical representation of his skills, then compared that profile with occupation descriptions and job postings that used different language. A person who wrote “changed tooling” could be matched with a posting that asked for equipment setup, even when the phrases were not identical.

That semantic matching was the part a spreadsheet could not replace. It let the tool infer related work from ordinary descriptions, and it widened Marcus's options beyond jobs with his old title. It also made a confident leap from factory experience to physically demanding roles because the same signals that suggested troubleshooting and production knowledge often appeared in warehouse, maintenance and supervisory postings.

An explanation panel showed that pay had substantial weight. The tool favored jobs near Marcus's target wage, followed by skill overlap and the number of openings within commuting distance. Physical demands were less dependable because employers described them unevenly, sometimes in a structured field the system could compare and sometimes in a paragraph near the bottom of a posting.

Marcus had not told the chat about his back. The tool treated the missing fact as no restriction, rather than as an unknown that should limit its confidence.

He added a line to the notebook: “Good wage, cannot do job.” By the end of the first week, that note appeared beside six recommendations.

A human career coach reviewed the same ranked list during an appointment. The coach could remove a job after Marcus explained why it would not work, but the next batch came from the same profile and returned to similar roles. The model had learned a strong pattern from his wage goal and work history. One rejected posting did not erase that pattern.

Marcus did not blame the coach. The list made sense if his body was left out of it.

The chat included a place to add preferences and barriers. It invited users to enter anything that might affect the search, which could include transportation, scheduling or physical limits. A notice said information could be available to authorized program staff and service providers that operated the tool. It did not tell Marcus, in language he found clear, whether a future employer would ever receive the text he entered or whether it would be used to improve later recommendations.

He was willing to describe what he could lift. He did not want to enter a diagnosis, treatment history or the name of a medication into a system that was built to find jobs rather than hold a medical record.

For two days, the box stayed empty.

The detail that changed the list

Marcus first tested the restriction in a temporary chat rather than saving it to his profile. He wrote that he needed work with no lifting above 25 pounds and no repeated bending. He did not name the injury.

The ranking changed at once. Quality documentation work moved up. So did production planning and dispatching. Warehouse jobs fell lower, although they did not disappear, and one inspection job still required lifting 40 pounds because that detail sat in the posting text rather than the structured physical-demand data the tool used more consistently.

The result showed both the value and the limit of the system. Once given a functional constraint, it could recalculate across hundreds of occupations faster than Marcus and his coach could review them one at a time. It could not reliably determine whether every employer's description was complete, and its polished explanations did not signal which underlying fields were missing.

Marcus copied four new recommendations into the notebook. Beside production planner, he wrote $25.60 and “mostly computer.” Beside quality records coordinator, he wrote $24.

90 and “ask about boxes.” The notebook had become a check on the ranking, not a copy of it.

He showed the temporary results to his coach and disclosed the 25-pound limit, without giving the diagnosis. The coach added the functional restriction to the part of his program record used for job matching. Marcus was told program staff could see it. He still did not receive a firm answer about whether the tool's outside operator retained the chat text separately, so he stopped using the open conversation box for health details.

The saved restriction narrowed the list. His estimated pay range also dropped. Jobs near $27 an hour were more likely to involve equipment repair, floor supervision or material handling, while desk-based manufacturing roles often started between $23 and $26 in his area.

That was information he could use, even though he did not like it.

Marcus chose a 16-week training program covering production scheduling software and quality documentation. The state program paid the tuition. During training, he used the career tool to translate his factory tasks into terms that appeared in office-based manufacturing jobs, then checked every recommendation against the employer's posting and his notebook.

Eight months after the layoff, he accepted a production scheduling job at $24.80 an hour. The role involved walking through the plant to check work in progress, but materials staff handled the lifting. He earned $2.

30 less per hour than before and had a clearer path to work he could keep doing.

He continued to use the tool after starting the job. It was good at finding adjacent skills he had not named himself, particularly the link between production logs and scheduling records. He never decided that the privacy notice was clear enough. His permanent profile contains “25-pound lifting limit.

” The diagnosis remains on paper in a folder at home.

The notebook is still beside his laptop. The first pages are filled with jobs paying close to $27 and lifting requirements of 40 or 50 pounds. On the later pages, the wage figures get smaller while the word “possible” appears more often.

Questions people ask

Can an

AI career tool account for a lifting restriction?

It can if the restriction is entered in a field or sentence the system uses for ranking. Marcus's tool lowered physically demanding jobs after it received his 25-pound limit, but incomplete job-posting data still allowed one unsuitable role through. The generated explanation sounded certain even when the physical-demand field was missing.

Why did the tool prioritize pay over physical safety?

Marcus entered a wage target, and the system had structured pay estimates it could compare across jobs. His physical limit was absent, while lifting demands appeared inconsistently in employer postings. The ranking therefore had a strong wage signal and a weak safety signal, which pushed well-paid factory and warehouse work toward the top.

Did

Marcus have to enter his diagnosis to improve the matches?

In his case, no. The useful detail was functional: no lifting above 25 pounds and no repeated bending. His coach recorded that limit without adding the diagnosis. The tool produced a better list because it could rank jobs against a concrete constraint, not because it knew the medical reason behind it.

Could an employer see the health information he entered?

The notice Marcus read identified program staff and service providers, but it did not give him a clear answer about later employer access or separate retention of chat text. He kept his diagnosis out of the system and saved only the work limit used for matching: 25 pounds.