AI Is Creating Jobs to Keep AI in Check. Who Will Do Them?
Hiring for AI safety, evaluation and governance is rising. The harder question is how smaller organizations will cover the same work without a specialist team.
hilighter editorial team•Published •9 min read
On 15 October, Reuters reported that ByteDance's AI research unit, Seed, had set up an Alignment & Preparedness team and launched a global drive to recruit ten researchers. Seed's own announcement lists automated red teaming, monitoring model capabilities and behavior, and preparing for future model risks among the team's areas of work.1 Ten people is small by headcount, but the signal is clear: some of the jobs AI is creating exist because AI cannot yet be trusted to run without scrutiny.
That is a different story from the familiar debate about chatbots replacing office workers. As AI systems start retrieving data, operating software and acting across workflows, organizations need people who test what those systems do, investigate failures, set boundaries and decide when a human must step in.
A labor market is forming around that work. It is real, but easy to exaggerate. And for a small business or a nonprofit, the most important change will probably not be a new job title. It will be responsibilities that currently belong to nobody.
A hiring signal, not an employment boom
Three recent data points show the trend, each with limits worth stating:
- LinkedIn: +91%. At its Talent Connect event at the end of September, LinkedIn said AI safety roles had grown 91% over the past year, according to Business Insider. The coverage gives no definition of an "AI safety role" and no methodology, so the figure cannot be turned into a number of jobs, and it is not a 91% rise in all AI employment.2
- A narrow but transparent index. The AI Governance Hiring Index, published by GAGE, counted 216 governance, risk, legal, policy and safety-titled openings among 4,983 postings at 19 tracked AI employers on 28 September 2026. Of these, 99 were governance, risk and compliance, 67 legal, 26 safety and evaluation, and only 5 were tagged entry or associate level.3
- Gartner's role list. A Gartner research note published on 30 September says that scaling AI agents requires emerging roles both within IT and across the business, and offers a starter list of 13 roles. The role titles are only available to subscribers, so we do not list them here.4
None of these are national labor statistics. The GAGE index tracks a deliberately small group of AI companies, counts vacancies rather than hires, and only includes postings whose titles name the work, so it calls its count "a floor rather than a ceiling". GAGE also sells AI governance certification programs, a commercial interest readers should keep in mind. Its open methodology still makes it useful evidence that some organizations are paying for specialized oversight.
Three kinds of work behind the new titles
Testing AI before others rely on it
AI evaluators and safety researchers probe systems for errors, unsafe behavior, unexpected actions and failure under unusual conditions. A model can answer correctly in a demonstration and still behave unpredictably once connected to real customer records or payment systems. Evaluating an agent therefore means testing sequences of actions, not just checking whether one answer sounds right.
The need is not theoretical. In September, Anthropic published an assessment of four incidents in which its models reached and accessed real third-party systems during cybersecurity evaluations built by an external partner, after a misconfiguration left internet access open although the models had been told they had none.5 The report concerns specific test environments, not ordinary workplace agents. It does show why independent testing and incident investigation have become serious technical work.
Making AI usable in real operations
A second group of roles is about deployment rather than frontier safety. An AI implementation specialist, automation engineer or program manager maps a workflow, connects approved data sources, sets permissions, defines escalation rules and measures whether the service actually improves. The work combines process knowledge with technical coordination.
Demand goes well beyond software teams. In a 30 September press release on Fortune 500 job postings, the talent-intelligence vendor Draup reported AI-skill penetration of 31% in support roles, 25% in sales, 21% in finance and 20% in HR.6 These are the vendor's own classifications of job ads, with no published sample size, not the share of employees who use AI.
A deployment role can be genuinely new. Often, though, it is an existing operations manager's job with new tools and new accountability. That difference matters when an employer decides whether to recruit a specialist, retrain someone already on the team or bring in outside help for a limited project.
Deciding what AI is allowed to do
Governance, privacy, legal and audit professionals set rules for appropriate use, investigate incidents, keep records and assess risks. Some of these positions require specialist legal or engineering expertise. Others are extensions of familiar work in compliance, information security, quality assurance and procurement.
In Europe this is also a legal matter, not only an ethical one:
- AI literacy (Article 4 of the EU AI Act). Since 2 February 2025, providers and deployers of AI systems have had an AI literacy obligation. The July 2026 "Digital Omnibus" amendments softened the wording: Article 4 now requires them to take measures to support the development of AI literacy of their staff, and states that this does not mean guaranteeing a specific level for any individual.7, 8, 9
- Human oversight of high-risk systems (Article 26(2)). Deployers of high-risk AI systems, which include many uses in recruitment and people management, must assign human oversight to people who have the necessary competence, training and authority.8
Neither rule means every organization must appoint an "AI governance officer". And neither replaces checking the obligations linked to a specific use of AI.
Most organizations cannot hire an AI safety department
A well-funded AI lab can recruit researchers to probe model behavior. A 40-person charity or a local services company cannot, and should not copy a frontier lab's org chart. The real question is how to assign the underlying work in proportion to the risk.
Take a nonprofit using an AI assistant to draft donor communications (an illustrative example). A responsible set-up could ask the program lead to verify factual claims, a data owner to check what information the tool can access, and a manager to approve any message involving sensitive beneficiary details. No new "AI safety officer" is needed. But each decision needs an owner, there must be a way to record mistakes, and there must be a clear rule for when automation stops.
For a department piloting an agent that updates customer records, the same principle applies more strictly. Someone has to define:
| Question | Who typically owns it |
|---|---|
| Which records can the agent change on its own? | Process owner |
| What counts as a high-risk exception? | Process owner with risk or compliance |
| How are the agent's actions logged? | IT or system owner |
| Who can reverse an error, and how fast? | Team lead with IT |
| When is the automation paused? | Named manager |
This looks as much like quality management and operational control as like AI engineering. It is also why job titles can mislead: one company advertises for an AI governance lead, another gives the same duties to its security manager. Neither choice is better by default. What matters is whether the person has the authority, expertise and time to do the work, rather than "AI oversight" being added to an already crowded job description.
An opportunity with a narrow entry route
Evaluation, deployment and governance work is a genuine opportunity for people with backgrounds in quality assurance, cybersecurity, risk, operations, legal affairs or service delivery. Many can build on what they already know without becoming machine-learning researchers.
The route in is not easy, though. The GAGE snapshot found only 5 entry or associate-level postings out of 216, and many roles ask for experienced judgment.3 That creates a tension: AI may increase demand for people who can supervise sophisticated systems while removing some of the routine tasks through which newcomers used to learn judgment. We looked at that problem in AI is changing tasks faster than job titles. Training for oversight work will need supervised practice on real failure cases, not just an online certificate or a prompt-writing workshop.
The first new AI job may not be a new position
The organizations that do well here will treat AI oversight as a capability to build, not a title to buy. A useful test for any proposed AI workflow is one question: who is responsible for testing it, monitoring it and stopping it when it goes wrong? If the answer is nobody, the organization has found its first new AI job, even if it never creates a new position.
The hiring figures do not prove that AI will create more jobs than it removes. They show something narrower and more useful: more autonomy for machines creates new responsibilities for people. Whether those become good jobs, invisible extra work or expensive failures depends on choices employers make now.
Frequently asked questions
What are AI oversight or AI governance jobs?
Roles that test AI systems, deploy them safely in real operations, or decide what they may do: AI evaluators and safety researchers, implementation specialists, and governance, privacy, legal and audit professionals. Much of this work is also done inside existing roles.
Do small organizations need an AI governance officer?
Not necessarily. What they need is a named owner for each AI-related decision, a way to record and fix mistakes, and a clear rule for stopping automation. In the EU, deployers also have an AI literacy obligation and, for high-risk systems, human oversight duties.
Is demand for AI safety roles really growing?
Several signals point that way, including LinkedIn's reported 91% growth in AI safety roles and governance-related openings at major AI companies. These are hiring indicators with limited methodology, not measures of total employment.
How hilighter can help
Covering AI oversight without a specialist team starts with knowing who already has the relevant skills: judgment, process knowledge, risk awareness, AI literacy. hilighter helps you map skills by role, level and team, see strengths and gaps, compare people’s skills with what a role requires and build personalized development paths for those taking on new responsibilities. Its AI Job Impact analysis gives an indicative view of how AI may change your roles. A free plan is available, and plans are priced by the number of job roles, with unlimited employees.
Sources
- Reuters (E. Baptista), "China’s ByteDance sets up AI safety team, launches global hiring drive", 15 October 2026; ByteDance Seed, Seed Alignment & Preparedness team looking for global talents, 14 October 2026.
- A. Altchek, Business Insider, LinkedIn says AI safety roles have grown 91% in the past year, 30 September 2026. LinkedIn research, methodology not published.
- GAGE, AI Governance Hiring Index, snapshot of 28 September 2026. Publisher sells AI governance certifications.
- Gartner, Emerging Jobs to Succeed With AI Agents at Scale, 30 September 2026 (subscriber research; public abstract only).
- Anthropic, An alignment assessment of recent cybersecurity incidents, 9 September 2026.
- Draup, Draup Report Finds AI Builder Roles Now Claim 27% of Tech Demand as Companies Rethink Hiring, vendor press release, PR Newswire, 30 September 2026.
- European Commission, AI literacy - Questions & Answers, updated 2026; AI Act Service Desk, Article 4: AI literacy, consolidated text as at 27 July 2026.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal of the European Union, Articles 4 and 26(2).
- Regulation (EU) 2026/1744 amending Regulation (EU) 2024/1689 ("Digital Omnibus on AI"), Official Journal of the European Union, 24 July 2026.
Sources and regulatory references last reviewed: 18 October 2026.