In a striking contradiction that underscores growing concerns about artificial intelligence in the workplace, researchers at Google DeepMind who specialise in AI safety are explicitly telling job seekers not to trust their own employer's hiring technology. The team managing applications for open positions has instructed candidates to complete a supplementary form alongside their standard application, specifically to circumvent the risk of being filtered out by Google's internal automated screening systems. This internal guidance, marked with warnings against sharing widely, reveals a troubling gap between the tools technology giants market to corporate clients and what those companies actually believe about their own systems' reliability.
Google aggressively promotes its artificial intelligence hiring solutions to corporate customers as a transformative way to rapidly process vast stacks of job applications and identify the most suitable candidates. The pitch centres on efficiency and speed—allowing human resources departments to handle volume at scale. Yet internally, Google DeepMind's AGI Safety and Alignment Team, tasked with researching how to reduce risks from powerful artificial intelligence systems, operates under the assumption that this very technology cannot be trusted. The disconnect highlights a broader industry pattern where vendors sell confidence in their AI systems while quietly harbouring doubts about their functionality.
The internal document acknowledges the problem plainly: there exists a meaningful probability that qualified candidates will be incorrectly screened out or that their applications will languish unreviewed within the automated system. Rather than wait for algorithmic processing, the team directs applicants to submit a specially designed form that ensures a team member will manually examine their credentials. This workaround essentially admits that the company's own vaunted efficiency tools create unacceptable risks for recruiting talent—a striking statement from one of the world's leading artificial intelligence development centres.
When confronted with this contradiction, a Google DeepMind spokesperson denied that the company's systems filter applicants incorrectly, yet simultaneously confirmed that the team established the special form specifically to bypass recruiter review. The explanation that "there are no shortcuts to getting hired" rings hollow given that the form itself functions as a shortcut around the automated system. This defensive response illustrates how technology companies often frame genuine technical limitations as deliberate design choices, preserving the appearance of confidence even when internal practices suggest otherwise.
The broader adoption of artificial intelligence in hiring reflects an industry-wide confidence in automation that may not be warranted. Companies deploy various approaches—some use AI models to rank candidates on a scoring system, others scan resumes for specific keywords and credentials. These systems ostensibly save time and introduce objectivity, yet the opacity of their decision-making and the technical challenges in building reliable systems create new problems. Human resources teams worry about their own obsolescence even as evidence mounts that these tools may be fundamentally flawed.
Google's Workspace division, which sells productivity software like Google Drive to businesses, actively markets new artificial intelligence capabilities aimed at revolutionising human resources work. The pitch promises to "save HR time by quickly creating drafts for job postings, evaluating resumes, and forecasting hiring needs." This product line exemplifies how technology companies monetise automation, positioning AI as the inevitable future of corporate hiring. Yet the hesitation within Google's own research divisions suggests the reality remains considerably messier than marketing materials acknowledge.
The reputational risk to AI hiring systems has intensified following investigative journalism and legal challenges that expose potential discrimination. A Bloomberg investigation discovered that OpenAI's ChatGPT exhibited signs of bias correlated with applicants' names, suggesting that artificial intelligence systems trained on historical data perpetuate existing employment discrimination. Meanwhile, Workday Inc, a major vendor of workplace management software, faces litigation alleging that its hiring systems screen out candidates based on race, age, and disability—violations of employment law. Though Workday maintains that humans make final hiring decisions and has denied wrongdoing, the lawsuit points to how opaque algorithmic systems can encode prejudice in ways difficult to detect or remedy.
The stakes are particularly high in Southeast Asia, where rapid tech adoption often outpaces regulatory frameworks. Malaysian and regional companies increasingly implement AI hiring tools without fully understanding their limitations or potential for discrimination. The cautionary example of Google's internal practices suggests that even the most technologically sophisticated organisations struggle to build hiring systems that reliably work as intended. Smaller firms adopting these technologies may lack the resources to identify and correct such failures, potentially disadvantaging entire applicant pools based on flawed algorithms.
Job seekers themselves have begun gaming these systems, using artificial intelligence to generate applications at unprecedented scale or to craft responses calibrated to pass algorithmic screening. Google DeepMind's form instructions specifically caution against this approach, noting that hiring teams "get really tired of reading LLM answers, because they all sound very samey." This arms race between applicants using AI to optimise their presentations and systems using AI to filter applications creates a recursive problem: artificial intelligence designed to identify the best candidates increasingly struggles to distinguish genuine human expertise from machine-generated text.
The tension between what Google sells to customers and what it practices internally reveals a critical failure in how technology companies approach accountability. When a leading artificial intelligence research organisation acknowledges that its own hiring filters are unreliable enough to require workarounds, it signals a more fundamental truth: the current generation of AI hiring tools operates beyond our ability to fully understand, predict, or control their behaviour. For Malaysian companies and regional organisations considering similar technologies, the lesson is clear—scepticism and human oversight remain essential, not optional.
This gap also raises questions about vendor responsibility and disclosure. Should companies marketing hiring AI systems be required to disclose known failure rates or limitations? If Google's own researchers cannot confidently rely on these tools, what standard should apply to customers with fewer resources for independent evaluation? The current regulatory vacuum allows companies to simultaneously market artificial intelligence as transformative while quietly building backup systems for when the technology fails, a pattern that benefits vendors while shifting risk onto employers and jobseekers alike.
