The New AI Bargain: Retrain Before Replacing

A city can sign a software contract in a season. A person whose tasks move does not gain a new competence on the same calendar. That gap is what a tentative Oakland agreement tries to name, in two sentences, before anyone can claim the jobs have been saved.

The sentences are narrower than a headline that says workers have already won. They are also more precise than a press release that only says "AI usage protections."

What the signed page actually requires

On 31 October 2026 the City of Oakland announced a tentative three-year agreement with the International Federation of Professional and Technical Engineers Local 21. The release lists a 9 percent wage increase over three years: 2.5 percent in January 2027, 2.5 percent in January 2028 and 4 percent in January 2029. It also lists transit benefits, equity pay for lower-paid classifications, and "AI usage protections." It does not print the AI text. It says the union would schedule a ratification vote and that the City Council still had to approve the contract. It says the union rescinded a strike notice for 5 November. The release does not say anyone has been retrained.¹

The AI wording is in the union's tentative-agreement packet, a scanned PDF linked from the Oakland chapter page. A cover index dated 5 November 2026 lists item 17 as passed and signed on 30 October 2026: an amended counterproposal on artificial intelligence. The clause page is headed 30 October 2026. Allyson Cook, chief spokesperson for the city, and Debra Grabelle, chief spokesperson for IFPTE, signed it. The handwritten dates read 30 October 2026 on the city line and 31 October 2026 on the union line. This article read that page from the image. It is a tentative agreement inside a packet. It is not a council ordinance.²

Section 13.9 says the city agrees to comply with all California law on the use of artificial intelligence by employers in the workplace, including but not limited to SB 947, SB 951, AB 1331 and AB 1883, each marked 2025-2026.

Section 13.9.1 is titled "Preservation of Bargaining Unit Work." It has two sentences. The city shall not implement or utilize AI systems that eliminate bargaining unit classifications. If AI systems reduce the number of positions needed for work traditionally performed by represented employees, the city will retrain and/or reassign them.²

A classification is not a headcount. The first sentence can be met while the number of positions falls, as long as the classification itself is not eliminated. The second sentence is the remedy the text gives for that reduction: retrain and/or reassign. It does not define hours of training, a test of competence, a suitable vacancy, or what happens if neither option exists. It does not say "rather than firing." That phrase is The Guardian's summary, in a report published on 12 November 2026. The direction is a fair reading. The words are not the clause.³

The Guardian also reports that on 22 October the union voted to authorize a strike, with 94.1% in favor, and that the 5 November strike was put on hold. The city's release says the notice was rescinded. The union's ratification page says the agreement was reached without having to strike. This article did not open a separate tally sheet. The 94.1% figure stays a newspaper report.¹³⁴

On 14 November 2026 that ratification page still called the deal tentative. The bargaining team recommends a yes vote. The vote is in person on 17, 18 and 19 November. Nothing on that page, and nothing in the 31 October city release, shows that members have ratified the contract or that the council has approved it. A signature by two spokespersons is not a lived practice. No source read for this article shows that a represented employee has already been retrained under section 13.9.1. The clause does not bind a private employer in Oakland, or a city in another state.¹⁴

The bills are named, not reprinted

The four numbers are the worker package listed in Governor Gavin Newsom's signing announcement of 30 September 2026. The subject lines on that page are: SB 947, employment and automated decision systems; SB 951, employment, technological displacement and notice; AB 1331, workplace surveillance; AB 1883, workplace surveillance tools. Lorena Gonzalez, president of the California Federation of Labor Unions, is identified there as speaking on those four bills. This article does not quote her, because the published sentence is not needed to identify the package.⁵

The same announcement summarizes the package in three workplace lines: employers would be prohibited from relying only on AI when making a disciplinary or termination decision; employers would have to be more transparent when a mass layoff, relocation or termination is caused by an AI system; and surveillance tools would be limited, including a ban on using them in a workplace bathroom. One published sentence is incomplete. It says "requiring employers to provide if" a layoff is caused by an AI system. This article does not treat that broken line as statutory text. The chaptered code was not opened in this pass, so effective dates are not stated here. A promise to "comply with all California law" does not, by itself, tell a manager which section applies to a tool that only drafts a letter.⁵

Pennsylvania asked an earlier, different question

On 21 March 2025 Governor Josh Shapiro wrote to Commonwealth employees. The letter sets out principles. Three of them matter beside Oakland. The administration says it will adopt generative AI for employee efficiency and productivity, not with the agenda of substituting Commonwealth employees. Where human discretion is involved in performance and oversight, generative AI will not replace human supervision, and it should not be a substitute for direct human supervision. A Generative AI Labor and Management Collaboration Group is to take input from employees, including people covered by a union. The letter also points training at the Office of Administration. The published lines use a dash and, in places, a contraction, so this is a paraphrase. The letter is a statement of purpose. It does not say "retrain and/or reassign" if the number of positions falls.⁶

The Philadelphia Inquirer, on 10 August 2026, reports a separate agreement with SEIU Local 668, which it says represents 10,000 state workers. Steve Catanese, the local president, told the paper the state committed to a process in which human workers oversee, review and guide generative AI outputs. He also said: "Is it fully guaranteed forever? No." His remark that Pennsylvania is, as far as he knows, the only state government with collectively bargained AI protections is his knowledge. It is not a census of contracts. This article did not open the SEIU text.⁷

Use is a different count from protection. On 15 April 2026 the Pennsylvania Office of Administration said more than 3,000 employees across 35 agencies were using generative AI, and that an additional 6,500 were enrolled in training required before use. The Inquirer, on 10 August, says more than 3,000 are using the tools, up from 175 in the pilot, and that 6,000 more were enrolled in training. The April enrollment figure and the August enrollment figure are not the same number. This article does not average them.⁷⁸

The April release also describes the pilot those 175 people ran, across 14 agencies. Participants reported an average of 95 minutes saved per day while using the tool to draft, summarize, research and brainstorm, with human review still described as essential. The Inquirer adds that 85% of participants called the experience somewhat or very positive. Those are participant reports. They are not an independent time study, and they are not evidence that jobs were preserved. The same administration page says unemployment chatbots do not invent answers and escalate when a question falls outside approved content, and that hiring guidance bars generative AI from making hiring decisions or answering interview questions. Those are design rules on a government page. They are not section 13.9.1.⁷⁸

A leader survey is a third kind of number. EY's findings page of 10 November 2026, on a survey of 200 US state and local decision-makers commissioned by the EY Center for Government Modernization, says 69% report full integration or implementation of generative AI. The Guardian's "nearly 70%" of agencies, and a July-to-August field window, are not the wording on that findings page. A self-report of implementation is not a count of retraining clauses. Reported benefits on the same page are also self-reports. This article does not use them as proof that services improved.³⁹

What a small team can write down first

The steps below are an editorial suggestion for an organization of about 40 people. Oakland did not agree to them. Pennsylvania did not test them. They are not a research result.

Start from tasks, not from a list of titles to delete. Write what the tool will actually do, what a person must still judge, and where a wrong output would harm a client, a colleague or someone who will struggle to contest an error. Name one accountable person for each workflow.

Then write the transition rule before the pilot, in words the budget can carry. One version: before an AI-related redundancy decision, look for a revised role and offer a defined period of paid training. That sentence is an example. It is not a quotation from section 13.9.1. If there is no real vacancy and no paid time, a promise to retrain is a slogan. Who had paid time, as distinct from a software license, is a separate problem. See why a training gap becomes a career gap.

Ask for evidence of the new work. Sitting through a webinar is not a redesigned job. A person can show a check: a generated letter compared with the rule, a case that should be escalated, or information that must not be pasted into a public tool. A manager can look at accuracy, the explanation, privacy and the decision to escalate. That list is a suggestion, not a validated test.

Give people a route to challenge a result when a tool touches scheduling, performance, hiring or discipline. Name the human reviewer and keep a record. The longer argument is who oversees the algorithmic manager. This article does not repeat it.

A counterexample belongs here. If a system saves ten hours a week and the training takes forty hours, the employer still has to separate a one-time cost from a recurring gain. If the tool only moves mistakes from an intake desk to a specialist, the hours described as saved may be harder hours somewhere else. Those figures are an illustration. They are not a time study.

There is also a record, already on this site, of what happens after a job is gone. In earlier waves of automation, the next search took longer, and retraining in the data was an association, not a guaranteed landing. See why the next search takes longer. A clause that says "retrain and/or reassign" is a process. It is not that outcome. A large employer can also announce cuts while a union asks for retraining, without a public count of who moved. See what the Barclays figures do not prove.

A European employer is not bound by Oakland

This article is not legal advice. A tentative California contract does not apply the EU AI Act, and the Act does not copy section 13.9.1.

Article 4, in the text the Commission service desk shows as amended, says providers and deployers shall take measures to support the AI literacy of their staff and of other people who operate systems on their behalf. They take account of technical knowledge, experience, education and training, the context of use, and the people or groups on whom the systems are used. The same paragraph says this obligation does not require a guaranteed level of literacy for any individual. The service desk says the wording reflects the EUR-Lex consolidated text as at 27 July 2026, including Digital Omnibus amendments, and that the summary is not legally binding. A literacy measure is not a duty to retrain or reassign before a redundancy.¹⁰

Annex III, point 4(b), lists AI systems intended to make decisions affecting the terms of a work relationship, promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance and behaviour. A drafting aid is not on that list merely because it saves time. A tool intended to score people for dismissal, or to assign work from personal traits, is a different question. Classification depends on the intended purpose. The service desk says Annex III rules apply from 2 December 2027.¹¹

Article 26 applies when a system is high-risk. Paragraph 7 says that, before putting a high-risk system into service or using it at the workplace, employers shall inform workers' representatives and the affected workers. Paragraph 11 says that deployers of Annex III systems that make or assist decisions related to natural persons shall inform those persons. Neither paragraph is a retrain-or-reassign clause. Neither is a finding about Oakland or Pennsylvania.¹²

The question worth keeping

The useful question is not whether a job title will vanish. It is who carries the cost when the tasks inside that title change, and what concrete route the person has to move with the work. Oakland's two sentences try to answer that before the reduction. On 14 November 2026 they were still a signed tentative clause, not a completed retraining. Pennsylvania's letter answers with a purpose, beside a pilot whose time savings are what participants reported. A smaller organization can write a shorter rule. It should not pretend the rule is already a statute, or already a result.

How hilighter can help

hilighter does not bargain city contracts, decide redundancies, or place someone in a vacancy. It can make the work under a role visible before a tool changes the tasks. Teams can map skills by role, compare people with those requirements in skills and talent insights, record a check or a correction in skills evidence and recognition, and set personalized growth paths that do not depend on whoever attended an optional webinar. A manager's review of that evidence sits in manager assessments. The AI job impact view is indicative. It is not section 13.9.1, and it is not a prediction that a role will disappear. The Free plan covers up to three job roles. The number of employees is not capped.

Sources

1. City of Oakland, Oakland Reaches Tentative 3-Year Agreement with IFPTE Local 21, 31 October 2026. Tentative three-year contract. Wage increase of 9 percent over three years (2.5 percent in January 2027, 2.5 percent in January 2028, 4 percent in January 2029), transit benefits, equity adjustments, and "AI usage protections," without the clause text. Union ratification and City Council approval still required. Strike notice for 5 November rescinded. No statement that anyone has been retrained.
2. IFPTE Local 21, City of Oakland Local 21 Tentative Agreement (July 1, 2026 - June 30, 2029), linked from the Oakland chapter page. Scanned packet. Cover index dated 5 November 2026, item 17 passed and signed on 30 October 2026. Section 13.9 and section 13.9.1 read from the page image on 14 November 2026. Header date 30 October 2026. Signatures: Allyson Cook, chief spokesperson for the city, handwritten 30 October 2026; Debra Grabelle, chief spokesperson for IFPTE, handwritten 31 October 2026. The two operative sentences are quoted in the article. Not a council ordinance and not evidence of completed retraining.
3. Ramin Skibba, The Guardian, US public workers are fighting to protect their jobs from AI - and some are winning, 12 November 2026. Strike authorization on 22 October, 94.1% in favor. The paper's summary says the tentative agreement requires the city to retrain or reassign workers displaced by AI rather than firing them. That "rather than firing" line is the paper's wording, not section 13.9.1. The 5 November strike described as put on hold. The page also says nearly 70% of state and local agencies have adopted or are implementing generative AI, citing an Ernst and Young survey from July to August 2026. Those adoption words and field dates are not on the EY findings page opened for source 9, so they are not used as the EY finding.
4. IFPTE Local 21, Oakland Ratification Vote, read on 14 November 2026. Tentative three-year agreement reached without a strike. Bargaining team recommends a yes vote. In-person vote on 17, 18 and 19 November 2026. Not a ratification result.
5. Office of the Governor of California, signing announcement, 30 September 2026. Subject lines: SB 947, employment and automated decision systems; SB 951, employment, technological displacement and notice; AB 1331, workplace surveillance; AB 1883, workplace surveillance tools. Lorena Gonzalez identified as speaking on those four bills; her remarks are not quoted. Summary lines on sole reliance for discipline or termination, transparency when a layoff is caused by an AI system, and surveillance limits, including bathrooms. The transparency sentence is incomplete in the published text ("provide if") and is not treated as statutory language. Chaptered code was not opened. Effective dates are not stated.
6. Commonwealth of Pennsylvania, Governor Josh Shapiro letter on AI and employees, 21 March 2025. Paraphrase of the purpose (efficiency, not an agenda of substituting employees), human supervision where discretion is involved in performance and oversight, the Generative AI Labor and Management Collaboration Group, and training through the Office of Administration. The letter uses a dash and contractions, so it is not quoted in full. It does not contain the Oakland "retrain and/or reassign" sentence.
7. Ethan Young, The Philadelphia Inquirer, As Pennsylvania implements AI tools for government employees, efficiency is the goal, not job cuts, 10 August 2026. SEIU Local 668 described as representing 10,000 state workers. Separate agreement, human-in-the-loop wording as reported. Catanese: "Is it fully guaranteed forever? No." His "only state" remark is his knowledge, not a census. More than 3,000 users, up from 175 in the pilot; 6,000 more enrolled in training, which differs from the 6,500 in source 8. Participant reports: 95 minutes a day and 85% somewhat or very positive. This article did not open the SEIU agreement.
8. Pennsylvania Office of Administration, Shapiro Administration Expands Safe and Responsible AI Across State Government, 15 April 2026. More than 3,000 employees across 35 agencies. Additional 6,500 enrolled in training required for use. Pilot: 175 employees, 14 agencies, average 95 minutes saved per day as reported by participants, on drafting, summarizing, research and brainstorming, with human review described as essential. Chatbots that match approved answers and escalate otherwise. Hiring guidance that limits generative AI to drafting, practicing and summarizing, and prohibits using it to make hiring decisions or answer interview questions. The page says the month marked one year since the collaboration group was created, with SEIU Local 668 employees and Department of Human Services management. Not evidence that jobs were preserved.
9. EY, 2026 State and Local Government Modernization Survey Findings, findings page dated 10 November 2026. Survey of 200 US state and local decision-makers, commissioned by the EY Center for Government Modernization. 69% report full integration or implementation of generative AI. Self-reported benefits are not used as findings here. Field dates printed by The Guardian were not on this page.
10. European Commission, AI Act Service Desk, Article 4: AI literacy. Wording reflects the EUR-Lex consolidated version as at 27 July 2026, with Digital Omnibus amendments. The summary is not legally binding. Paragraph 1 does not require a guaranteed level of literacy for any individual, and it is not a retrain-or-reassign duty.
11. European Commission, AI Act Service Desk, Annex III, point 4(b): decisions on terms of work, promotion or termination; task allocation based on individual behaviour or personal traits; monitoring and evaluating performance and behaviour. When does enforcement start?: Annex III rules from 2 December 2027.
12. European Commission, AI Act Service Desk, Article 26, paragraphs 7 and 11. Information duties for employers and for people subject to Annex III high-risk systems that make or assist decisions. Conditional on a high-risk system. Not a retrain-or-reassign clause.

Sources last reviewed: 14 November 2026. The Oakland AI clause was read from the scanned tentative-agreement page, not from the city's press release. The Guardian's "rather than firing" line is not the clause. Member ratification was still scheduled for 17 to 19 November 2026, and council approval was still outstanding in the 31 October release. The chaptered text of the four California bills was not opened, so effective dates are not stated. The SEIU agreement was not opened. The 95 minutes and the 85% figure are participant reports. EY field dates in the Guardian were not on the findings page. EU dates follow the Commission service desk and are not legal advice. Neither an AI-detector score nor the absence of Google penalties can be guaranteed.