OK.

‘No creativity, please, no wordart, or graphics, don’t stand out, unreadable.’ 

OK.

‘Don’t clutter your CV with photos, tables, columns, headers, or footers, unreadable.’ 

OK.

How many of these little nuggets have you read or heard about when the topic of résumés comes up? I heard them all and many, many more – from ‘experts’ and fellow job seekers alike. Everyone wants that edge, but not too edgy. Clarity, but using a specific format. Showcasing you, but … there’s always a but.

While researching this topic, I had my résumé open the whole time to see if it stood the test. It never did. I’d rephrase and rewrite sections according to what I read. Only to revert back or make new changes. Each article gives hope of the ‘perfect formula’ to pass the Applicant Tracking System (ATS) and get the attention of the human on the other side.

But here’s where it gets interesting – there’s no unified ATS that all companies use. No. Each company uses a different system, with its own unique algorithms. And you, dear applicant,  need to figure out which ATS the company uses so you can tailor your résumé to pass it.

How about that for a plot twist.

ATS in Action

The way ATS works is simple: it reads résumés by electronically analyzing (parsing) the relevant information, like name, education, and experience, then sorts them for the recruiter in a searchable format. 

One of its common features is keyword search – the system looks for specific words already mentioned in the job description, e.g. a specific number of years’ experience, certain skills, or a location. It also tracks candidates through the whole process, from application to interview results, and saves their information even if they were not selected, for future opportunities.

Beyond screening, it also handles job postings, interview scheduling, compliance reporting, and candidate notifications – managing the entire hiring workflow from start to finish. The system is built to handle large numbers of résumés, which is why it’s widely used to streamline the hiring process and free up time and resources.

A Timeline

The hiring process was not as sophisticated as it is now. Companies sent job postings to newspapers, you read the ad and sent your résumé by post or in person, recruiters waded through the lot and chose the most suitable candidates and invited them – by mail or by phone – for an interview. 

Too slow. Something had to be done to speed things up. Enter ATS. 

In the 1970s, it was just basic data entry with limited reporting capabilities. The 1980s saw added features like résumé parsing for faster sorting and analysis. The drawback was that it was expensive and difficult to use, which made it only implemented by large enterprises.

With the emergence of the internet in the 1990s, job postings and applications moved online. The system saw more advanced algorithms like candidate evaluation and ranking.

From the 2010s onwards, the Cloud enabled smaller companies to use ATS due to scalable and flexible subscription payments. The system’s analytics and reporting capabilities became more advanced, tracking criteria like cost per hire and time to fill. As mobile technology evolved, more and more candidates began using their mobile devices and social media accounts to apply.

When it comes to integrating AI, ATS is no exception – many providers have already built it into their products. There are even predictions that the future might even bring virtual reality and augmented reality technologies that could change the interview process completely.

ATS as a Business

The business of ATS software is booming. 

A 2025 report shows that the global ATS market was worth $2.5 billion in 2024 – a 12.3% jump from 2023. The top ten vendors alone controlled over 50% of that market, led by iCIMS at 10%, followed by Oracle, Workday, and Greenhouse Software. Annual growth among the top vendors was sharp: in one year, Workday grew 15.3% and Greenhouse 13.2%.

Reading the profiles of these leading vendors, one thing is very clear – AI is the star of their products. iCIMS has expanded its AI footprint with conversational tools and assessment features, even piloting autonomous AI Agents to handle sourcing and interview tasks. 

Oracle integrates generative AI into its Recruiting Cloud, using embedded co-pilots for job description optimization and candidate ranking. Workday is building a multi-agent ecosystem through its Illuminate system. Greenhouse is evolving into an agent-oriented platform, enabling third-party conversational AI to autonomously screen and schedule interviews.

Together, these plans paint a picture of ATS companies positioning AI as the backbone of the entire hiring process – shifting routine tasks to automation, with more on the way.

A 2026 report by 360 Research Reports puts the market at nearly $5 billion and is projected to reach $13 billion by 2035. As of 2025, 94% of Fortune 500 companies use an ATS, and 78% of large enterprises have one built into their hiring process. Small and medium businesses are not far behind – 62% have adopted cloud-based platforms. AI-powered screening tools grew 46% year-over-year, and demand for mobile-friendly and analytics-driven systems has grown 59% since 2022. In the new analysis, Workday now holds 14% of global market share, while Oracle holds 12%.

Two things stood out in the 360 report key findings: a ‘56% rise in AI-driven candidate scoring and predictive analytics tools’, and ‘47% of new product launches focusing on AI, machine learning, and mobile optimization.’

What this means is that the AI-powered ATS train didn’t just leave the station, it’s almost at its destination. The demand for it is growing faster than you can say ‘bias.’ Companies implementing it will not look back now, even with four out of ten – according to the same report – saying they struggle with integrating and migrating data from older HR systems. 

AI in hiring is here to stay.

Built on Broken Data

The study by Douglas Guilbeault and colleagues reveals how generative AI like ChatGPT perpetuates gender and age bias in hiring. When prompted to create over 34,500 résumés for 54 jobs using typical male or female names, ChatGPT portrayed women as younger and with less work experience than men.

When asked to evaluate the same résumés, the AI ranked older men highest in quality, putting older women and younger applicants at a disadvantage – the same groups that already face discrimination in the real world. This happens because the model draws from internet data filled with stereotypes (e.g., men are better at ‘fixing things’ and therefore suited for roles like construction) – amplifying societal biases rather than reflecting objective reality.

Guilbeault observes that AI companies are aware of the problem. But their fix is to add filters to block the most obviously biased outputs. He argues that this barely scratches the surface – it misses subtler biases like the age and gender gaps the study found. Real progress means tackling bias at the core of how these models are built, not patching after the fact. Until then, his advice is simple: be cautious. These tools can make you believe the issue of bias is resolved when it’s really not.

Another study tested several LLMs from OpenAI, Google, Anthropic, and Meta by having them score over 360,000 randomized résumés with different gender and racial identities. Compared with equally qualified White men, most models gave higher scores to female candidates (both Black and White) but lower scores to Black male candidates. These differences translate into real hiring impacts: women would have a higher chance of being selected, while Black men faced reduced odds.

Together, the two studies show that AI bias in hiring cannot be solved with patchwork fixes. It shows up in how résumés are generated, how they’re scored, and whose careers pay the price. With these inherent biases baked into the data, why do we expect AI-powered ATS to be fair in résumé screening?

Gaming the System

Writing a résumé should be an easy task – you put in the basic information an employer needs to consider you for a job, right? Wrong. It’s more than that, much more. I thought I knew a thing or two – I’ve had one for years. But after a career break, getting back into the job market, writing my résumé was anything but easy.

There is so much advice out there on how to write one, so I won’t repeat it here. What I want to focus on is something most of that advice misses – it’s not the usual suspects. Many websites advertise ATS-friendly templates complete with checker and scorer services, and if it doesn’t meet the minimum score, you need to rewrite it. Fine, I can live with that. What I’m not fine with is having to look for which ATS the company is using to tailor my résumé according to its algorithms.

For a simple experiment, I tried a LinkedIn job search, clicked on the ‘Apply’ button – not ‘Easy Apply’ – and it took me to the original website for the job. The platform the job was sourced from uses iCIMS. It was easy to see, I found it at the bottom of the page: ‘Powered by iCIMS.’ It’s also in the URL: ‘nameofplatform.icims.com/jobs/’

‘What do I do with that info?’
Easy, you google ‘how to optimize résumés for iCIMS’, which will give you many articles that try to explain how that specific system works. For example, it prefers simple fonts – Arial, Calibri, Times New Roman – and verbatim keywords from the job posting. This doesn’t just work for iCIMS, but for Workday, Oracle, and others. And if you can’t find the ATS yourself, Jobscan – for a fee – can do that with simple steps and give you optimization tips as well.

Is that something you needed to know? Absolutely, yes.
Will it make job searching easier for you? A resounding no.

Being a minimalist, I thought since my résumé has the basics, what more would an employer need? But I was mistaken. It’s the ATS I need to get past to reach the actual human. I hope you think of this new information as intel, not extra work.

But sometimes job candidates take it too far. 

I recently read about how Amazon is banning job seekers from using AI tools during interviews. Given the volume of candidates it receives, the company issued a set of guidelines to its internal recruiters to create ‘a fair and transparent recruitment process.’ Unless explicitly permitted, applicants may be disqualified if they used AI tools during the interview.

The problem became so widespread that Amazon shared tips on common signs the applicant is using an AI tool. Sometimes candidates sound as if they are reading instead of speaking naturally, even correcting themselves when they misread a word. Their eyes may follow text or drift away rather than focusing on the conversation. They might give confident answers that don’t directly address the question, or appear distracted and confused when reacting to AI‑generated outputs that don’t make sense.

And honestly? That’s only fair. You already have a foot in the door – showing your true self and answering naturally is the right way to go.

That said, and as a fellow traveler down the same road, I completely understand. Each step – learning about the job, tailoring your résumé, waiting for a response, and finally getting the ‘we invite you for an interview’ email – all that takes its toll, and you want the edge, any edge to ace that interview.

But if you’re considering using AI at that specific stage, think about it. What if it’s just a chat to tell you about the company and field your questions? What if it’s a work test? Who’ll be doing the actual work if not you? 

What if the company uses a platform like HireVue? Instead of a standard interview, candidates complete short game-based tasks designed to measure things like pattern recognition, working memory, and problem-solving. You could be tasked with performing actual job scenarios to test whether you can do the work, not just talk about it. What then?

The Human Cost

When hiring systems shut out qualified people, the economic consequences go far beyond the individual. Historically marginalized communities, including women and people with ethnic backgrounds, end up missing out on jobs – meaning fewer chances to grow in a career, build stability, or move toward long‑term security. 

Over time, this kind of exclusion widens existing wealth gaps and keeps certain groups stuck in cycles of underemployment and limited opportunity. And because well‑paid jobs often come with benefits like health insurance and retirement plans, being pushed out of those deepens inequalities even further.

Disclaimer: The following might be sensitive for some. If you find that it resonates with you to the point of disrupting your daily life, please, seek professional help.

Taking the consequences a step inward to what it does to the person on the receiving end – a psychological phenomenon called ‘job rejection fatigue.’ It’s the ‘emotional and mental exhaustion that builds up from receiving repeated job rejections over time.’ This doesn’t happen from one email – it’s the compound effect of several disappointments. It affects not just your confidence but also your health and social relations. 

But how does it compare to the stress that naturally comes from job searching? The simple answer: they’re two different beasts. When you’re looking for a job, you’re evaluating everything – from the role itself, to how many applicants, to your fit, to the company, etc. The uncertainty of the process causes stress. Job rejection fatigue is very specific to those rejection emails. Every single one piles it on till you reach a point when you fear opening the email to read the verdict.

The common advice you hear in this situation is ‘don’t take it personally.’ It doesn’t work and here’s why. The phenomenon is deeply connected to evolution: when the early human was rejected from the tribe/group/clan, it meant potential death. Those who took it ‘seriously’ survived, and you’re their descendant. So, it’s only an instinctive response – feeling the weight of it the way you do.

Here are the common signs to watch out for:

Emotional Symptoms: decreased motivation to apply, anxiety before opening emails, doubting your qualifications or career choices.

Behavioural Changes: reduced application quality, delaying search tasks, avoiding networking events, or withdrawing from social gatherings.

Physical Symptoms: changes in sleep patterns or in appetite, increased headaches, muscle tension, or fatigue even after getting rest.

If ignored, it might affect your life in the long run:

Relationship Strain: feeling irritable, withdrawn, volatile, or negative about your prospects can affect the people in your circle.

Career Stagnation: ‘settling’ for a role that doesn’t align with your qualifications or goals, and accepting terms without negotiating for salary or benefits.

Long-Term Confidence Issues: continued anxiety even after landing a job, imposter syndrome – doubting your abilities even when you’ve clearly earned your place – not pursuing better opportunities, nor building connections that might lead to better prospects.

If caught early, job rejection fatigue is manageable, here’s how:

Start with your mindset: think of every ‘no’ not as a judgment on who you are, but as a sign that it was the wrong fit. 

Change your tactics: instead of exhausting yourself with many applications, focus on a few thoughtful ones. This way you stay connected to your intentions and goals. 

Practice mindfulness: have small rituals to process disappointment – take a walk or do breathing exercises, just set aside some time for your body to recover. 

Take stock of how far you’ve come: be it skills, clarity, or resilience – you haven’t been standing still. It’s just a dry spell not a failure.

You need to always remind yourself that there’s more to you than a job seeker – you have your routines, your people, your hobbies, the parts of your life that shine despite someone else’s decision. Through all of it, give yourself grace; so much of hiring is outside your control, and the effort you’re making is already evidence of what you’re made of.

Fighting Back

One of the most famous cases is Derek Mobley v. Workday, Inc. 

In 2023, Mobley sued Workday alleging its AI tools discriminated against him based on race, age, and disability. The case was first dismissed because Workday was classified as an employment agency, but was later accepted in 2024 after a federal judge ruled that the company had a role in the decision‑making process by using workforce data from its customers’ companies to train its AI without accounting for the bias already present in that data.

In May 2025, a California district judge certified the case as a collective action suit – meaning anyone affected by the platform’s AI decisions could join. By July 2025, the case expanded to include individuals affected by HiredScore, an AI tool used by Workday customers to score, sort, rank, and screen applicants. The court also ordered the company to produce a list of their customers.

A more recent case was filed in California by job applicants Erin Kistler and Sruti Bhaumik against Eightfold – an AI-hiring platform – in January 2026. The platform works by assessing applicants and predicting whether they’d be a ‘good fit’ for a job based on résumé and job listing data.

According to the lawsuit, Eightfold builds profiles on job seekers that go beyond listing their qualifications. They assign personality labels, such as ‘team player’, rank the quality of their education, and predict where their career is headed. It’s accused of doing so by collecting this data without the applicants’ knowledge, consent, or the chance to correct any mistakes.

Kistler and Bhaumik use an already existing law – the Fair Credit Reporting Act (FCRA). They argue that Eightfold’s profiles function just like credit reports: sensitive data, collected and used to make decisions about people’s futures. FCRA makes sure that sensitive information collected by credit bureaus and similar agencies is only shared with people who have a legitimate reason to see it. It also requires companies to investigate disputes and to tell consumers if a credit, insurance, or job decision goes against them because of what’s in their report.

At the time of writing, a ruling hasn’t been issued for either case.

For many, lawsuits are a last resort – who wants the stress, the hassle, and the cost of a long litigation process? But sometimes it’s the only way to get any justice at all. 

If successful, those involved could get some closure. More importantly, job hiring platforms would be pushed to put their house in order – auditing their tools and cleaning up biased training data before it hurts someone’s future. 

‘…, There’s a Way’

In 2021, Dr. Sandra Wachter, professor of technology and regulation at the University of Oxford’s Internet Institute, developed a bias test – the Conditional Demographic Disparity (CDD). It’s designed to reveal when one group is rejected more often than it is accepted.

Here’s what it would look like in hiring: we have equal numbers of men and women applying for the same job. Now think of their applications falling into two piles – the rejected pile and the hired pile. If women keep landing in the rejected pile more than the hired pile, while the opposite happens to men, something is clearly wrong. That imbalance is what the test flags as demographic disparity. The group losing out is marked as ‘disfavored,’ while the group benefiting is marked as ‘favored.’ The data simply speaks for itself.

A strong example of how effective this test can be comes from a 2024 audit in the Netherlands. Investigators found that the Education Executive Agency (DUO) was unfairly flagging students with non‑European migration backgrounds for extra checks, leading to indirect discrimination. The issue was serious enough that it was sent to the Dutch Parliament, and the minister for Education, Culture and Science issued a formal apology once the findings were published.

Wachter’s CDD has already proven it works. The real question is why it hasn’t been widely adopted – it’s in companies’ best interest to fill positions with suitable candidates the first time around, without waiting for a lawsuit to force the issue.

Another solution comes from a study where participants were given AI-generated résumés with White, Asian, Black, or Latino-sounding names or other indicators of race, and asked to recommend who should be invited for an interview.

The results show that when people made hiring choices alongside AI recommendations, they often mirrored the system’s biases. With fair suggestions, participants chose fairly – but when the AI displayed moderate bias, people followed it almost completely. Even under severe bias, they still followed the AI’s lead about 90% of the time.

The researchers concluded that human decision-makers tend to trust AI guidance unless the bias is very obvious, and propose a solution: the implicit association test – a psychological tool used to reveal hidden or subconscious biases people may not realize they have. According to the study, biased decisions dropped by 13% after taking it. They recommended adding such training alongside educating recruiters about AI’s limits as a way to reduce hiring bias.

There are also recommendations for companies to better streamline hiring without compromising transparency or efficiency. The frustration of job seekers who spend six to twelve months looking might make some turn to litigation, which could be incentive enough for hiring teams to take staff training for better practices more seriously.

Companies should optimize their tools for fairness and not just efficiency by reevaluating historical data the AIs are trained on – it might be perpetuating bias. The assessment and predictive tools meant to test the applicant’s ‘cultural fit’ with a prospective employer might discriminate against certain demographic groups.

With the EU AI Act and California’s SB 53 (Transparency in Frontier AI Act) treating the use of AI in hiring as a ‘high‑risk’ activity, companies must now exercise stricter oversight throughout the process and meet compliance requirements – otherwise they risk facing severe legal consequences.

The bottom line for companies: audit current screening processes, check whether AI is independently making decisions, and ensure each step is fair and job-related. The goal is a system that catches and mitigates inevitable human mistakes – perfect hiring doesn’t exist.

When I read about these methods and recommendations for the first time, I put myself in the shoes of employers and thought implementing them might be difficult, expensive, or time-consuming. But then again, do companies really thrive on never-ending hiring rounds, turning away the best person for the job, negative reviews, or liability lawsuits? 

If not … 

Your move.

What’s ahead

Some problems just don’t have an easy fix – AI-powered ATS is shaping up to be one of them. The signs are there, the effects are becoming more visible every day, and the consequences are easy to predict. ‘Where there’s a will, there’s a way.’ It’s the will that is missing.

In some cultures including my own, work is not just about having an income – it’s tied to the self-respect and social standing that come from an honest day’s work. Being unemployed and actively looking should not be an extra burden on top of the demand of making ends meet. It cannot turn into a game of charades: I’ll pretend to apply to a human, a human will pretend to hire, and AI is right in the middle making all the decisions.

How about making life easier for everyone by dropping all pretenses. No one is buying it anymore. All those webinars, podcasts, articles, and advice might be helpful if the system were open and its decisions understood. Not knowing why you were rejected is frustrating enough. Getting a rejection email that lists possible reasons and you pick which one applies – that’s not transparency, that’s busywork.

Looking for an edge in the job market should come from your qualifications, potential, talents, skills, and experience, not from fixing a résumé with the right fonts and keywords so an AI can read it – that’s not building a career, that’s choosing wallpaper.

Let it stop with us. 

Make the next person sending out applications feel that they’ll be evaluated for their achievements. That the process is fair and the decisions explainable. That the barriers are not invisible. That the system actually works.

The post Applicant Tracking Systems: The AI That Broke Hiring appeared first on Untold.

Imagine this: you’re sitting at home watching the news. An employee in a major beverage company is coming out of the courthouse in an ongoing, much publicized legal battle with her employers. All she did was inform government officials of a specific component that has been added to a recent product without proper tests. After going through the proper channels and filing a complaint at the company, she was told that it would be investigated, but nothing happened. 

She took action. Because someone had to.

The company argues that she signed a non-disclosure agreement (NDA) that prevents any employee from divulging trade secrets. Her lawyers counter that when it comes to public health, NDAs are not legally (nor ethically) binding. 

You have been following this case from the start, and you’re waiting for the court-ordered independent lab results to settle the matter. 

Now, ask yourself this: which part of that scenario did you care about most? You’d only be human if the first thing that came to mind was personal interest. When it is a health-related topic, everyone’s first priority is – and should be –  their own. The second thing might be a coin-toss between how the court would penalize the company if wrongdoing was proven, and how government officials would try to revise laws on food additives.

Coming last would be the employee who reported it. Why? Because we all assume she’ll be fine.  She actually did the right thing and her actions will save lives. If the courts rule in her favor, she’ll be lauded as a heroine. She presented evidence: lab results, testimonies, and her own eyewitness account. What more is there to think about?

She’ll be fine.

And you’d be right to think that way. In the United States, whistleblowers in the food industry are protected under the Food and Drug Administration (FDA) Food Safety Modernization Act (FSMA). The company can’t retaliate against her in any way for whistleblowing: by dismissal, demotion, or transfer. Should that happen, she’d be within her legal right to file a complaint, win it, and likely be compensated and reinstated back to her position.

Let’s use the same scenario but shift the industry; this time it’s an AI company. Here, the story takes a different arc. 

The common theme among AI whistleblowers differs from that of the food and beverage industry. In the absence of clear laws and reporting channels, the decision to speak out against AI industry giants is often weighed against livelihood, personal and professional reputation, and, in one case, life itself.

Stories from the AI Trenches

Sounding the alarm takes not just evidence, but courage and a firm moral belief that you’re doing the right thing: telling the truth. 

Meet a few brave individuals who faced hard obstacles to let the world know what goes on behind closed lab doors.

Leopold Aschenbrenner: Former OpenAI safety researcher, fired from his position for allegedly sharing documents externally. He warned of “egregiously insufficient” security against foreign threats, citing OpenAI’s failure to adequately protect critical algorithmic information and model weights. He became an advocate, urging a shift away from fast, unsafe deployment towards robust safety measures.

Suchir Balaji: Former researcher at OpenAI and perhaps the most tragic story on this list. He resigned in August 2024 after stating the company’s use of copyrighted material violated U.S. law and posed commercial harm to creators. He was set to testify in intellectual property lawsuits against OpenAI; but was tragically found dead by suicide on November 26, 2024. 

Timnit Gebru: Co-lead of Google’s Ethical AI team, dismissed in 2020 after writing a research paper that exposed how current AI training methods could deepen biases against minorities and marginalized communities. Her dismissal gained widespread coverage, exposing corporate retaliation, prompting her to found the Distributed AI Research Institute (DAIR).

Louis Hunt: Former CFO and VP of Business Development at Liquid AI. He resigned from his position and publicly challenged the claim that AI models don’t replicate copyrighted works. He  presented evidence of generated outputs that were exact copies of texts from The New York Times, Stephen King’s books, and Harvard Business Publishing articles.

Margaret Mitchell: Founder and co-lead of Google AI ethics unit who was fired in 2021 for alleged misconduct. She testified in the September 2024 Senate hearing on AI oversight, stressing the need for clear instructions to employees navigating NDAs, and accessible whistleblowing channels to provide support to those who wish to come forward with their concerns.

William Saunders: Ex-OpenAI technical staff member who testified at the September 2024 Senate hearing. He advised on: 1) establishing a list of government contacts who understood the reported issues and could act on them, and 2) identifying legal protections insiders need when flagging actions that don’t break laws, but put public safety at risk.

Helen Toner: Former OpenAI board member on the nonprofit arm, who testified at the September 2024 Senate hearing. She highlighted how vagueness in current AI whistleblower laws discourages people from coming forward, specifically those with complaints about AI development that often don’t fit existing legal categories designed for traditional industries.

Anonymous and Named Whistleblowers: Group of eleven-thirteen current and former employees at leading AI companies who in June 2024 wrote an open letter “Right to Warn” calling for principles to create a safe environment for employees to voice their concerns on potential risks. In July 2024, they filed a complaint to the Securities and Exchange Commission (SEC) requesting an investigation into how NDAs restrict scrutiny of safety behaviors (e.g., the rushed testing of GPT-4o). 

This list is not exhaustive, many more names can be mentioned, including, David Evan Harris, Jacob Hilton, Geoffrey Hinton, Daniel Kokotajlo, Ramana Kumar, Jan Leike, Neel Nanda, Carroll Wainwright, and Daniel Ziegler, among others. 

Why Blow the Whistle

Merriam-Webster defines a whistleblower as “an employee who brings wrongdoing by an employer or by other employees to the attention of a government or law enforcement agency,” followed by the following note: “A whistleblower is commonly protected legally from retaliation.” 

If we analyze the stories above based on that definition, we see events unfolding in two stages. Stage One: the employee notices wrongdoing and first reports it internally. When the response from the company dismisses the complaint or addresses it inadequately, the employee either reports their concerns to the government/the media or publishes their findings as independent research.

The complaints themselves span a wide range, from ethical and governance concerns to more technical issues of alignment and safety. In addition, the alarms raised aren’t about one company’s specific attitude towards AI training, testing, and deployment; they bring to light an industry-wide, systemic pattern of behavior that if left unchecked could lead to non-reversable consequences.

For an insider to have that first-hand experience and courage to go through the steps to prevent such dangers, their actions should be celebrated as a reflection of moral integrity, not cause for retaliation. This brings us to the second part of the definition where the key term is “commonly protected”. Commonly means usually/often and that is an apt description in the case of AI whistleblowers. 

Stage Two is where we fail them, because there is simply no protection. The stories show that all of them lost their jobs, either by dismissal (with or without stated reasons) or through resignation (as a form of protest). They became industry pariahs just for speaking up, and some couldn’t find employment immediately after their stories broke out. 

If you’re an employee in an AI frontier company and read about the consequences faced by those who actually wanted to help, would you risk your career/livelihood, your industry standing, and your future to report AI risks? 

Nuclear Whistleblowing as a Model

In recent years, AI research has seen a surge in academic papers, books, interviews, and podcasts that wrestle with growing safety concerns. These works highlight AI’s dual-use and emphasize its black box nature – the opaque decision-making processes that continue to baffle even the scientists building these systems. 

In response, many AI governance and ethics experts compare the risks posed by advanced AI systems to those of the Chemical, Biological, Radiological, and Nuclear (CBRN) materials and agents in their severity and potential for catastrophic and even existential harm.

Therefore, as a model for the AI industry, I turn to the nuclear field. Its whistleblower laws and regulatory frameworks have long maintained vigilant oversight over both plant operations and worker safety.

The Energy Reorganization Act of 1974 (ERA) established the Nuclear Regulatory Commission (NRC), which regulates civilian (not military) nuclear facilities and materials. The Act was substantially strengthened in 1992 through the Comprehensive National Energy Policy Act, which added significant whistleblower protections including Section 211.

The key provisions in this amendment include the following:

1. It defines the protected activities for which employees cannot face discrimination. 
2. It establishes a clear complaint and investigation process for whistleblowers who face retaliation. 
3. It requires that these protections be posted permanently in the workplace for constant employee access
4. It directs the NRC or the Department of Energy (DOE) to conduct swift investigations into whistleblower allegations of substantial safety hazards.

Here are examples of how this works in practice:

In May 2013, the Occupational Safety and Health Administration (OSHA) found that Enercon Services Inc. wrongfully fired a senior engineer at Wolf Creek Generating Station for reporting safety violations. The engineer’s employment had been terminated in January 2012 after pointing out that soil coverage for buried safety pipes didn’t meet federal requirements. He was also asked to write a report justifying inadequate backfill material, but he refused.

OSHA ordered the company to reinstate the engineer with back pay, benefits, and compensatory damages. Enercon appealed, claiming the termination was for legitimate business reasons, but investigators found the engineer’s concerns were valid and that the field errors weren’t his fault.

In 2016, the Government Accountability Office (GAO) released a comprehensive review of DOE whistleblower protections and found that while the legal framework exists, enforcement and implementation needed strengthening. The report’s six recommendations, which were mostly accepted, demonstrate that Nuclear whistleblower protections are actively monitored, regularly evaluated, and continuously improved to address gaps, and ensure the law’s promise is realized.

In 2024, the U.S. Supreme Court ruled in Murray v. UBS Securities that whistleblowers under the Sarbanes-Oxley Act (SOX) don’t need to prove their employer intended to retaliate. The Second Circuit Court of Appeals had required plaintiffs to prove employer’s intent, but the Supreme Court vacated that decision. It held that if an employer treats someone worse (by firing, demoting, or changing their working conditions) because of whistleblowing (a protected activity), that constitutes a violation. The employer’s motivation is irrelevant.

This ruling is significant because the Supreme Court noted that the Energy Reorganization Act protections already worked this way. Under ERA, whistleblowers only need to show their protected activity was a factor in the adverse action. Then the employer must prove by clear and convincing evidence they would have taken the same action anyway. The Murray decision reinforced what nuclear whistleblowers already have: strong legal protections that shift the burden away from those who speak out.

The established path for Nuclear whistleblowers could serve as a blueprint for their AI colleagues, who operate in a similarly high-risk domain.

The Push for AI Whistleblower Protection: A Timeline

We now have a question: if AI risks are widely acknowledged as severe, from systemic bias to existential threat, why does the law still fail to protect those who report these risks? The answer is quite simple: existing legal frameworks typically require evidence of fraud or illegality – thresholds that may not encompass AI safety concerns. This leaves insiders vulnerable as they attempt to alert the public and policymakers. 

The following timeline traces key events from 2024 to 2025 that have shaped the need for whistleblower protection. 

In May 2024, news of OpenAI’s restrictive NDAs became known. Following public criticism, CEO Sam Altman admitted he was unaware of their extent and expressed embarrassment. He confirmed the company’s revisions of these agreements to remove provisions that threatened to deprive departing employees of their vested equity. This acknowledgment came after reports that OpenAI’s NDAs prohibited ex-employees from criticizing the company or disclosing safety concerns, sparking broader scrutiny of AI industry practices.

In June, thirteen current and former AI employees wrote an open letter “Right to Warn” urging frontier AI companies to promote an environment of safety-first in AI development and deployment. They stressed the current practice of forcing new hires to sign an NDA, the terms of which demand that they cannot voice concerns or disparage the company, even after leaving.

Moreover, they raised the point of widespread retaliation against whistleblowers, which adds to the growing concern that AI companies fear losing funding and investments more than protecting humanity from a technology that could possibly destroy it. By not allowing criticism, the companies silence well-meaning experts who could steer innovation in the right direction.

On July 1, the same group filed a formal complaint to SEC’s chairman Gary Gensler and to Senator Chuck Grassley’s office. In it, they provided evidence that OpenAI’s NDAs were restrictive to any protected disclosures of concerns related to AI safety, and asked the chairman to conduct an investigation into whether that practice broke SEC rules.

On July 31, OpenAI sent a letter to senators outlining robust safety measures, including dedicating 20% of computing resources to safety efforts like red-teaming and risk evaluations. The letter also affirmed support for whistleblower protections by introducing anonymous reporting options, for example: the Integrity Line.

On August 1, Sen. Grassley sent a letter to Sam Altman demanding answers about OpenAI’s restrictive NDAs with a deadline of August 15. The questions included:

1. Whether OpenAI changed the restrictive language of their NDAs, and provide proof of it.
2. The number of employees who requested to contact federal authorities, including all the relevant details.
3. The number of SEC investigations into OpenAI, including basis and outcome.

Sen. Grassley’s aim was to identify the purpose of the NDAs: whether it is for protecting trade secrets, or preventing employees from voicing their concerns.

On September 17, Sen. Blumenthal chaired a hearing by the Senate Committee on the Judiciary’s Subcommittee on Privacy, Technology, and the Law. He and several senators heard testimonies from, and directed questions to, expert witnesses on current AI regulations from an insider’s perspective. The hearing covered topics relevant to AI safety implementation and whistleblower protection.

At the start of 2025, President Trump signed Executive Order 14179 in January titled “Removing Barriers to American Leadership in Artificial Intelligence,” which revoked Biden’s AI safety order (EO 14110 of 2023). The new EO mandated the creation of an AI action plan within 180 days, and explicitly set the policy of maintaining U.S. “AI dominance” by removing regulatory barriers to innovation. This shift in policy marked a decisive turn towards deregulation, one that positioned acceleration as a national priority.

In May, Sen. Grassley and bipartisan cosponsors introduced the “AI Whistleblower Protection Act” (S. 1792) in response to mounting concerns about retaliation against AI employees who raise safety issues. The bill would prohibit retaliation against both employees and independent contractors who report AI security vulnerabilities or safety violations. As of October 2025, the bill has yet to advance beyond the Committee on Health, Education, Labor, and Pensions.

Following the executive order mandate, the White House released the AI Action Plan in July. It advances the deregulatory vision by explicitly prioritizing speed and innovation, while dismantling what it calls “onerous regulation” and “bureaucratic red tape.” By eliminating safety requirements, the plan effectively grants the private sector carte blanche to accelerate AI deployment with minimal oversight.

In October, California passed SB-53 “Transparency in Frontier AI Act”, the first U.S. state law establishing frontier AI transparency and whistleblower protections. Among other provisions, SB-53 introduces direct whistleblower protections for covered employees tasked with assessing or managing critical safety risks:

• It mandates that large AI developers establish anonymous reporting channels.
• Employees are shielded from retaliation when using these channels or when reporting externally to state or federal authorities.
• Protection applies when the whistleblower reasonably believes their employer’s actions pose a substantial threat to public health or safety; especially in cases involving catastrophic risk.

These provisions mark a shift from viewing whistleblowers as disruptors to recognizing them as vital protectors of public interest.

From corporate controversies to whistleblower advocacy and legislative breakthroughs, the road to stronger safeguards has been anything but easy, especially when the drive for rapid deployment outweighs the call for scrutiny. Yet through the combined efforts of courageous individuals and responsive institutions, robust protections are finally within reach.

Line of Defense

In her testimony at the September 17 2024 Senate hearing, Dr Mitchell compared AI training to baking:

“With Data as essentially ingredients, training is cooking, and the model is the output … we’re missing an approach where we have recipes, a deep understanding of what the pieces are that result in this output.”

In a landscape where the builders themselves don’t fully grasp the systems they create, insider disclosures may be the only path to effective regulation.

As the race to develop and deploy AI accelerates, safety and alignment concerns take a back seat at frontier companies. Governance frameworks are racing to catch up by creating laws and policies to protect public welfare, but the gap between innovation and governance remains wide. We need whistleblowers now more than ever. 

But how can we rely on insiders to shield us from high-stakes AI failures when they lack industry-wide protections and remain vulnerable to retaliation? Whistleblowers are our primary line of defense in this field. But shields need protection too. Without enforceable legal guarantees, we’re asking people to make sacrifices they shouldn’t have to.

Where is the justice in that?

The post Whistleblowing as a Shield: Protecting the Voices That Keep AI Safe  appeared first on Untold.