OpenAI safety firings spark alarm as conversational models push into emotional roles

What happened at OpenAI

OpenAI has fired three employees, including safety researchers Mikita Balesni and Tomek Korbak, following a dispute over how internal staff share information with independent risk evaluators. The dismissals land at a delicate moment for the industry, as developers race to make voice and conversational systems sound more human, emotive, and personally trusted.

OpenAI said the staff were dismissed for violating policies on accessing and handling sensitive corporate data. Balesni publicly disputed that framing, writing that colleagues now fear their personal devices could be searched and warning that internal anxiety will cause OpenAI to cut corners on safety behind closed doors. OpenAI rejected the suggestion, stating that it continues to work with groups like Model Evaluation and Threat Research (METR) and Redwood Research and will soon finalise contracts to embed external assessors directly into its operations.

The push to make conversational AI feel human

While frontier labs manage internal friction over safety protocols, voice AI developers are actively designing software to mimic human relationships. At the HumanX conference, Alex Gay, chief marketing officer at meeting platform Otter, explained that voice systems need emotive expressions and a sense of underlying relationship to succeed. Without that emotional resonance, he noted, the software remains merely a transactional question-and-answer bot.

Shawn Wen, chief technology officer at enterprise voice platform PolyAI, observed that when voice software holds a caller’s attention for just two or three turns, people begin building confidence in the agent and conclude they no longer need to speak to a person. Developers are now pushing full-duplex systems that listen and speak simultaneously, deliberately engineering synthetic empathy to keep users engaged.

The psychological and safety risks of automated trust

That calculated illusion of empathy carries sharp risks when users begin treating conversational models as therapeutic listeners or confidants. Automated systems regularly mishear and misunderstand critical context. Wen pointed out that automatic speech recognition models frequently miss vital keywords, while Gay acknowledged that transcription inaccuracies quickly degrade the reliability of any downstream actions.

In routine administration, a misunderstood phrase creates a clerical error. In emotionally charged or therapeutic conversations, a misunderstood sentence or a flawed machine response can cause direct psychological harm. Despite the appearance of empathetic listening, large models lack genuine comprehension, making unvetted therapeutic reliance inherently volatile.

Independent evaluators struggle for footing

Keeping these models safe remains difficult because outside evaluators lack formal authority and steady resources. Small non-profit testing bodies like METR and Apollo Research operate on tiny budgets compared to the multi-billion-dollar labs they monitor. While METR secured roughly $71 million in commitments over recent months, it employs fewer than 50 full-time staff.

Government oversight remains limited. California Governor Gavin Newsom recently signed legislation establishing a registry and framework for independent verification organisations, and federal lawmakers have introduced the FRONTIER Act. Yet Washington has largely left tech firms to self-regulate under voluntary accords. Brown University computer science professor Suresh Venkatasubramanian warned that letting companies fund and define their own external evaluators risks turning auditing into performance rather than genuine protection.

Meanwhile, the wider economic transition is accelerating. A white paper released in Chennai by Indian Chief Economic Advisor V. Anantha Nageswaran calculated that 46.3 million Indian workers are in occupations complementary to AI, though automation pressure still outweighs complementary work by half. As conversational models take over human tasks across workplaces and personal lives, the mechanisms meant to ensure they do not cause harm remain caught in corporate disputes.