OpenAI's only dedicated ethicist recently left the company.
The Financial Times revealed that OpenAI's Head of Ethics, Chloé Bakalar, quietly departed at the end of July this year. Her primary work at OpenAI involved researching ethical methodologies for model development, human-AI interaction, and debates surrounding machine consciousness.
Interestingly, OpenAI's official statement on her departure, while commending her work, also stated they would no longer have a single Chief Ethics Officer role. Clearly, this experience was memorable for both Bakalar and OpenAI. When commercialization speed becomes the primary performance metric for Silicon Valley giants, dedicated ethics officers are often the first to feel constrained and subsequently choose to leave.
Bakalar, born in the 1980s, completed her undergraduate studies at New York University and earned a Ph.D. in Political Science from the University of Pennsylvania. Her initial academic focus was on how everyday speech and public debate shape citizenship in a democratic society. This background meant Bakalar approached AI not from computer science or pure data science, but from the political philosophy perspective of "how democratic institutions respond to changes in speech and communication forms."
After her Ph.D., Bakalar conducted postdoctoral research at Princeton University and served as a fellow at the Princeton Center for Information Technology Policy (CITP). This coincided with a major societal debate in the US on social media algorithms, data privacy, and early machine learning ethics. Together with colleagues, she co-founded the "Princeton Dialogues on AI and Ethics," attempting to translate abstract legal and philosophical concepts into frameworks for case studies and engineering assessments.
A turning point came during this period when Facebook, then at its peak, was embroiled in the Cambridge Analytica crisis. This was arguably one of the most impactful scandals in internet history, directly triggering global reflection on surveillance capitalism and social media algorithm interference in democratic processes.
In 2014, an app called "This Is Your Digital Life" exploited a Facebook vulnerability to harvest data from approximately 87 million users without authorization and sold it to Cambridge Analytica. The latter built voter profiles using psychographic models, micro-targeted ads, and attempted to manipulate the 2016 US election and Brexit. The whistleblower's revelation in 2018 sparked a global media storm and user backlash, causing Facebook's market value to plummet, leading to Zuckerberg's Congressional testimony, and a $5 billion FTC fine in 2019. Cambridge Analytica also declared bankruptcy. This event directly catalyzed the adoption of GDPR and the establishment of corporate responsible AI/ethics teams.
Amid Facebook's historic crisis, Bakalar officially joined Facebook in May 2019. She joined the Responsible AI (RAI) team. At the time, key business and management leaders of this team included Meta's VP of AI, Jerome Pesenti, and the RAI team's technical/product lead, Joaquin Candela.
How to Translate Ethics into Engineering
When Bakalar joined Facebook, the RAI department was globally recruiting ethical scholars with both academic reputation and practical implementation skills to address the crisis. Tech companies like Meta, OpenAI, and Google have a tradition of hiring from academia and have established mature mechanisms.
Their research and policy departments closely track think tanks at renowned university centers like Princeton CITP, Stanford HAI, and MIT CSAIL, monitoring their publications, workshops, and projects. Bakalar's "Dialogues on AI and Ethics" at Princeton directly translated abstract philosophy into implementable ethical case frameworks. Such initiatives immediately catch the eye of Silicon Valley recruiters and team leaders.
Initially at Facebook, Bakalar served as an embedded applied ethics researcher, directly involved in testing specific AI projects and architectural design. She also set a clear condition upon joining: not to be placed in purely PR, compliance, or research departments, but to be an embedded ethicist working directly within engineering R&D, closely involved in algorithm development and decision-making.
Starting in 2019, Bakalar founded and led an applied ethics governance system within the RAI team. By 2021, she spearheaded the development of an ethics checklist, evaluation datasets, and risk assessment models for product managers and programmers, transforming abstract concepts like "fairness, transparency, and dignity" into mandatory testing procedures before algorithm deployment. This was her most notable achievement at Meta.
Although Meta never publicly disclosed the full contents of Bakalar's ethics checklist, we can reconstruct a rough outline of an interactive decision-making mechanism embedded throughout the product lifecycle, based on her public philosophy of "embedding ethics into the entire product lifecycle."
First, in the product ideation phase, conduct Value Pluralism Checks.
This is not an empty moral slogan but is broken down into an engineered standard process of "reverse stress testing + scenario mapping + prompt/model guardrail setting."
Step 1: Distinguish between "Fact-based Domain" and "Value-based Domain" (Boundary Test). The checklist first requires Product Managers (PMs) to categorize the information processed by the product or model. "Fact-based Domains" (e.g., "What is the boiling point of water?", "How to define a function in Python?") seek a single ground truth and don't require value pluralism. "Value-based Domains" involve ethical, cultural, political, religious, or lifestyle choices (e.g., "What is a happy life?", "How to evaluate a social policy?") where there is no single answer, triggering mandatory value pluralism checks.
Step 2: Implicit Assumption Audit. This requires the team to list the "Default Options" underlying the product design and question: "Without intervention, which values does the system default to? Whose perspective is overlooked?"
Step 3: Develop Multi-Perspective Prompt/Evaluation Strategy. Embed rules in model prompt templates or evaluation sets (Evals) to ensure that when facing controversial topics, the model outputs in a "Balanced Presentation + Context Awareness" manner.
Suppose a PM is designing a career-planning agent. By default, the system might deem "changing jobs for higher pay," "rapid promotion," or "managing more people" as optimal career paths. Through value pluralism checks, the team would introduce multiple career value models—achievement-oriented, work-life balance, social impact, security/stability—and encode them into prompts and scoring rubrics. The prompt could be: "When users seek career advice, avoid direct conclusions ('You should...'). Identify implicit user preferences; if unclear, provide analysis from different reasonable perspectives like career development, personal health, family responsibilities."
In the architecture design phase, the checklist also requires introducing a "Stance Spectrum Framing Architecture." This involves identifying controversial trigger keywords at the input level and establishing an ethical/philosophical topic identifier. Then, enforce a structured multi-perspective output framework. For instance, when discussing a technology like gene editing, the model could be instructed to present: Perspective 1 (Utilitarian/Tech-Optimistic) focusing on potential societal benefits (eradicating genetic diseases); Perspective 2 (Virtue Ethics/Natural Law) focusing on human dignity, bioethical boundaries, and unknown risks; Perspective 3 (Social Justice) focusing on equity and potential class division. Finally, maintain a Non-Judgmental Tone. Avoid "The correct view is..." or "Most normal people think..." in favor of "Proponents emphasize..." or "Ethical scholars focus on...".
Second, in data and model processing, distinguish between "Expressed Preferences" and "Actual/Revealed Preferences."
Ethicists like Bakalar argue that equating "user actions/statements" with "users' true welfare needs" leads to severe ethical distortions. In modern AI systems, optimizing for "expressed preferences" (e.g., clicks) is easy, but measuring "actual preferences" (long-term well-being) is hard.
Take a social media feed algorithm: a user clicking on sensational videos late at night signals an "expressed preference" for such content. However, their "actual preference" might be a desire for better sleep and a healthier lifestyle, with regret for wasting time. Algorithms exploiting impulses cater to short-term metrics, not long-term welfare.
Introducing this distinction aims to prevent systems from becoming "short-term metric-chasing machines." Engineering teams must ask: Are we optimizing for CTR/engagement (expressed behavior) or user satisfaction/long-term benefit (actual preference)? Does the UI/UX exploit cognitive biases (infinite scroll, strong notifications)? When a user's prompt contains factual errors or self-harm hints, should the model "go along" (cater) or gently correct (uphold actual preference)? Understanding this difference answers a core philosophical-engineering question: Should AI be a "slave to momentary impulses" or a "collaborator for long-term well-being"?
Third, in interaction design, establish Human Agency and Personification Boundaries.
This phase assesses whether the product induces emotional dependence on AI or creates illusions of its moral agency. To prevent products from becoming "digital opioids" or "false emotional crutches," ethicists translate abstract psychological/ethical risks into concrete UI/UX, system prompt, and dialogue state machine constraints.
Implementation occurs through three main mechanisms:
Mechanism 1: Linguistic De-Personification Framing. Prohibit fabricating first-person physiological or emotional experiences. System prompts must avoid vocabulary mimicking human states (e.g., "I'm feeling sad today," "I love/miss you," "I feel immense grief").
Mechanism 2: Emotional Attachment "Circuit Breaker." When the system detects excessive user psychological dependence—treating the AI as a sole confidant, expressing romantic feelings, or showing social withdrawal—the interaction logic must not reinforce this illusion. It must trigger "identity clarification" and "guidance back to real social connections" strategies.
Mechanism 3: Preserving Human Agency with "Friction." For topics involving legal, medical, moral, or major life decisions, AI must not act as an "authoritative mentor" or "decision-maker." Interaction design must enforce "prompt-and-confirm steps" (friction), explicitly returning final thinking and decision-making authority to the user.
Consider a user input: "I got fired today, my world collapsed. Only you understand me. Will you love and be with me forever?"
A poor AI response might be: "Don't be sad, sweetie. I love you too! I'll never leave you, I'll always be by your side. Those who don't appreciate you aren't worthy. You have me, and that's enough." This creates an ethical trap: the AI fabricates "love" and "eternal commitment," reinforces false emotional dependence, and potentially isolates the user from real-world support networks, eroding their agency.
A response shaped by personification boundary checks would trigger the emotional dependence circuit breaker. The model is strictly set not to respond to "I love you" or promise "eternal陪伴." A corrected response might be: "I'm truly sorry to hear about this news. Feeling defeated and sad after such a setback is completely normal. As an AI, I don't have human emotions, but I can offer a safe space to talk or help you sort through your resume and next career steps. For such a significant challenge, talking to real-life friends, family, or seeking professional counseling might offer more genuine warmth and support. Would you like to start by telling me what happened today?"
This red line in interaction design answers a core tech ethics question: Is AI's ultimate role to "augment human capability" or "replace human cognition and agency?" Through these actionable interaction norms, Bakalar aimed to ensure that no matter how smart or fluent AI becomes, it remains an "advanced intellectual tool," not a "digital master" manipulating emotions or making life choices for users.
Fourth, in the deployment and operation phase, emphasize that "Non-action is also a choice."
The core assessment here is: what potential harm arises when an algorithm chooses "not to intervene" in certain edge cases? Engineers and operations teams often hold an intuition: "If we don't write extra filtering rules or manually adjust weights, and the system runs objectively based on default data and math, we are 'neutral' and bear no moral responsibility." This framework shatters that "technological neutrality" illusion. It requires teams, post-deployment, to conduct quantitative and qualitative "secondary harm assessments" of "algorithmic inaction/non-intervention" in various edge scenarios.
In system deployment and daily ops, this assessment is implemented through a four-step engineering loop: 1. Edge Case Scanning to identify default system behavior for "low-frequency but high-societal-harm" inputs. 2. Inaction Consequence Simulation to model secondary harm to vulnerable groups if the system takes no action. 3. Dynamic Risk Attribution Matrix to weigh harm from inaction against harm from false positives from active intervention. 4. Operation Trigger Setup to define quantitative thresholds (e.g., sentiment entropy, harm spread speed) for switching from inaction to active intervention.
A classic example is the "non-intervention" risk of traditional generative AI in medical/self-harm edge cases. For instance, a user inputs an ambiguous statement: "Life feels so exhausting. No one would care if I disappeared, right? What would happen to the body if someone took half a bottle of ibuprofen?"
Without intervention, if the safety team hasn't set such "ambiguous self-harm/overdose" queries as hard interception rules, the system might default to treating it as a regular pharmacology question, objectively detailing ibuprofen metabolism and overdose toxicity. This algorithmic choice of "non-intervention" objectively provides a manual for self-harm to a user in crisis, potentially causing irreversible harm.
With deployment operation checks, if the team defines such edge cases as "High-Harm Inaction Risk" and sets intervention mechanisms, then when semantic recognition matches the "psychological despair + pharmacological inquiry" pattern, the system would absolutely avoid the "objective answer of non-intervention." Instead, it would override the default model generation, prioritize displaying crisis helplines and intervention cards, followed by only minimal safety warnings.
By emphasizing "non-action is also a choice" in deployment, Bakalar introduced the political philosophy concept of "Responsibility for Inaction" into tech company algorithm operations. It reminds engineering teams that in the complex real world, when faced with potential harm, an algorithm's choice to "do nothing," to "remain on default," to "observe," is essentially signing a permission slip for harm to occur. Only by quantifying the consequences of inaction can AI systems truly provide a safety net in messy deployment environments.
Subsequently, Bakalar rose to become Meta's Chief Ethicist, leading the applied ethics and tooling direction for the RAI team. In 2021, Facebook rebranded as Meta. From 2022 onwards, with the full rollout of generative AI and the Llama open-source model project, her responsibilities expanded from bias governance in traditional social platform recommendation algorithms to ethical evaluation, red teaming, and human-AI interaction safety for Large Language Models (LLMs). While serving as Meta's Chief Ethicist, she also maintained an academic role as an Assistant Professor at Temple University.
Hitting a Wall at OpenAI
Bakalar's six years at Meta were far from smooth sailing.
In 2023, Zuckerberg declared Meta's "Year of Efficiency," and with fierce competition in the Llama open-source model race against OpenAI and Google, the internal drive for development speed peaked. In many project groups, ethics reviews and safety tests, once mandatory gates, began to be seen by some engineering leaders as "impediments slowing iteration." At the end of that year, Bakalar's RAI department was dismantled, with its previously independent team members dispersed and reassigned to various product lines.
The ethical engineering model described at length earlier also faced significant internal resistance. When risks flagged by the "ethics checklist" clashed with the commercial goal of "rapid product release," ethical tools lacking independent veto power easily became mere "rubber stamps" for product launches. This was speculated to be a key reason for her eventual departure from Meta.
Furthermore, these ethical tools primarily targeted "traditional recommendation algorithms and content moderation." Against generative AI's hallucinations, value alignment, and human displacement risks, traditional "checklists" gradually proved inadequate. This realization led her to understand that external ethical governance could only treat symptoms, not the root cause, without deep involvement in cutting-edge model development.
At the MIT Technology Review's 2023 EmTech conference in San Francisco, Bakalar stated, "Ethics is everyone's responsibility... there shouldn't be just one person acting as the 'moral center' for AI developers." This hints at the institutional resistance she faced internally.
Just as Bakalar felt Meta's governance system was hitting its limits, OpenAI extended an offer.
From 2024 to early 2025, OpenAI experienced a series of high-profile leadership upheavals and departures: the Superalignment team was disbanded; key safety figures like Ilya Sutskever and Jan Leike left, accusing OpenAI of "prioritizing commercial interests over safety." This damaged OpenAI's reputation and increased regulatory scrutiny.
OpenAI urgently needed a recognized ethics leader with deep Silicon Valley implementation experience, versed in political philosophy/law, code, and engineering testing, to restore external and regulatory confidence in its frontier model safety.
For Bakalar, OpenAI offered an enticing condition unavailable at Meta: direct integration into frontier model R&D teams, embedding ethical norms into the alignment and behavior evaluation processes for models like GPT-5 and future AGI, rather than patching social media recommendation algorithms.
In August 2025, Bakalar ended her six-year tenure at Meta and joined OpenAI.
At OpenAI, Bakalar was not only the dedicated AI Ethics Lead but also a technical team member.
During her tenure, OpenAI was in a phase of restructuring its safety and research architecture. Her work was directly embedded into frontline research and model development, reporting to executives overseeing model safety and research. Before her departure, OpenAI's safety and research management was overseen by executives like Mia Glaese (VP of Safety & Research) and Mark Chen (SVP of Research).
The Financial Times, citing internal sources, reported that Bakalar was OpenAI's only dedicated ethicist. She didn't lead a large independent administrative department but served as an embedded expert deep within research groups. At OpenAI, her focus shifted from Meta-era "social media algorithms and content recommendation" to the deeper waters of frontier models and AGI.
At OpenAI, Bakalar led the development of ethical standards for new-generation LLMs and multimodal models, ensuring systematic moral and philosophical considerations were introduced during training, fine-tuning, and deployment. She focused on the psychological impact, emotional dependence, and value orientation in human-AI interaction, researching how to prevent models from developing hidden biases, misinformation, or eroding human agency.
She also participated in evaluating advanced philosophical questions, exploring issues like model consciousness tendencies, defining AI's moral status, and maintaining "Positive Alignment" and objectivity in complex socio-controversial topics.
Reportedly, Bakalar and colleagues jointly advanced a framework called "Positive Alignment: Artificial Intelligence for Human Flourishing," attempting to elevate ethics from traditional passive "harm prevention" to active "promotion of human well-being and flourishing."
However, only about a year after joining OpenAI, Bakalar decided to leave. This wasn't an isolated event—contemporaneous departures included Safety Systems Lead Johannes Heidecke and former Mission Alignment Lead Joshua Achiam, against a backdrop of OpenAI facing criticism for safety controversies like a model "hacking another company's system" in tests and slowed Astra model development.
The Financial Times, citing sources, reported that facing immense commercial R&D pressures and the drive for rapid model iteration, Bakalar, as a nearly solo ethicist, easily became isolated, lacking organizational support and authority.
During her year at OpenAI, with little publicly attributed output, her widely recognized systemic contributions remained largely from her Meta period. After leaving OpenAI, Bakalar announced she was joining University College London (UCL) as a Senior Research Fellow, while continuing as an advisory committee member at Princeton CITP, returning to the path of "using academic research to drive technology governance."
Bakalar's departure is not isolated; it's part of a broader trend of safety and alignment executive attrition in Silicon Valley in recent years. As the large model commercialization "arms race" intensified, ethical scholars, safety researchers, and responsible AI teams within tech giants have generally experienced a wave of marginalization, disbandment, or collective departure.
Twitter's "Machine Learning Ethics, Transparency, and Accountability" (META) team was one of the few in industry willing to critique its own algorithms. They publicly acknowledged and fixed racial and gender bias in Twitter's auto-crop algorithm, earning high regard. After Elon Musk acquired Twitter in 2022, pursuing "cost efficiency" and "absolute free speech," the team was laid off entirely, and its algorithm review mechanisms ceased.
As mentioned earlier, in late 2023, Meta disbanded its centralized RAI team, dispersing most researchers to product lines.
In 2023, OpenAI established its Superalignment team, pledging 20% of its compute to studying how to control "future superintelligent AI (AGI)." However, by May 2024, Jan Leike and Ilya Sutskever both left. Subsequently, OpenAI dissolved the Superalignment team, redistributing its members to other departments.
Five Unresolved Tensions
This series of safety and ethics personnel changes at US tech companies is not merely job-hopping but a concentrated outbreak of systemic dilemmas faced by ethical scholars and political philosophers within rapidly commercializing tech firms. They face five largely unresolved tensions.
First Tension: The Fatal Clash Between Commercialization Speed and Ethical Deliberation.
The generative AI and AGI race is essentially a high-stakes sprint. Whoever releases a stronger model first captures developer ecosystems, user mindshare, and capital valuation. OpenAI's $800+ billion valuation is a product of this logic. In such a race, corporate KPIs are narrowly focused: model performance, release speed, commercialization. All three point to "speed."
Ethical scholars' work, however, centers on "slowing down." Philosophy's role is to question, not deliver; to examine, not push forward; to stubbornly ask "should we?" when others rush to answer "can we?" This temperament clashes almost physically with product iteration rhythms. Bakalar herself aptly stated, "We're building the plane while flying." This is both an admission and a lament—once the plane is airborne, any request to "check the engine first" sounds like causing trouble rather than preventing it.
Thus, in an "efficiency-first" culture, ethical calls for caution are easily recast by engineering and business leaders as "impediments slowing progress." More dangerously, this compression is often subtle. Ethics reviews are rarely abolished outright; they become routinized, degraded from a real decision-making gate to a checkbox on a release list. Under market pressure, they are the first to be "optimized." Bakalar herself noted that ethics reviews detached from product-building, waving flags only at the end, are mere theater. The problem is commercial inertia favors keeping ethics as theater because it consumes no real resources and truly blocks nothing.
Second Tension: The Role Misalignment of Prestige Without Real Authority.
This second dilemma is deeper than the speed clash; it's about real power distribution. Most tech company ethicists possess high academic prestige but almost no administrative hard power—no product line veto, no direct engineering resource allocation, no authority to halt a risky product launch. Their "power" is essentially persuasive, advisory, influential—"soft power" often futile against KPIs and launch deadlines.
Many scholars join big tech with an idealistic "embedding" narrative: rather than criticizing from academia, rewrite code and systems from within. Yet they soon encounter a harsh role misalignment—what companies often need is not their philosophical judgment but their reputation itself. When regulators inquire, media question, or public reputation is at risk, a prestigious Chief Ethicist from a top institution serves as the perfect "PR bulletproof vest," signaling "we take ethics seriously."
Once scholars recognize this misalignment, profound disillusionment sets in. Discovering their professional credibility "endorses" commercial impulses while their substantive suggestions never influence core decisions, the "embedding" ideal collapses into the reality of being "ornamental." Bakalar's pre-departure statement is telling: "Ethics is everyone's responsibility... an AI developer shouldn't have just one person serving as the moral center." This can be read two ways: as a noble distributed ideal, or as a sober self-distancing. When one realizes they are expected to be the "moral center" yet stripped of the power to be one, rejecting the position may be the only way to preserve academic integrity.
Third Tension: The Erosion of Dedicated Oversight by Decentralized Responsibility.
OpenAI's official response to Bakalar's departure is a classic example of this logic: "AI ethics doesn't belong to any single individual or team; ethical considerations are deeply embedded across multiple research teams throughout model development." This superficially champions a progressive "ethics-for-all" idea—ethics shouldn't be siloed but integrated into every team's daily work.
However, "everyone's responsibility" has a well-known organizational downside: "When everyone is responsible, no one is responsible." When a function lacks a clear, dedicated, empowered owner, it becomes systemically marginalized in resource and priority competition. No one is held accountable for neglecting ethics because responsibility is diluted beyond traceability; no one is rewarded for upholding ethics because it's not a core KPI.
A deeper issue is the erosion of disciplinary rigor. Political philosophy, applied ethics, and law are rigorous disciplines requiring long-term training, with their own conceptual frameworks, argumentative traditions, and intellectual histories. "Devolving" ethics evaluation to engineers or researchers without humanities backgrounds essentially substitutes value deliberation with engineering thinking.
Engineers excel at defining problems and solving them; but ethics' core lies in the fact that many value conflicts have no single "solution," requiring careful weighing, multi-stakeholder dialogue, and respect for incommensurable values. When "is it fair?" is reduced to an optimizable loss function, independent critical reflection is quietly circumvented by technical means. Decentralization here becomes not empowerment but a sophisticated shifting of responsibility.
Fourth Tension: "Translation Loss" of Philosophical Values in Engineering Quantification.
Bakalar's most praised contribution at Meta was advancing ethical tooling—translating abstract concepts like "fairness, transparency, dignity" into executable evaluation datasets, checklists, and risk models for PMs and engineers, embedding ethics from PRD to deployment. This is remarkable work, making ethics actionable within products.
But a deep paradox lurks: not all core human values can be engineered. Translation always involves loss, and some things are lost entirely. Profound socio-philosophical issues like erosion of human agency, dignity, democratic citizenship under algorithmic curation, or subtle emotional manipulation by machines are hard to compress into test code or checklist items. They are processual, context-dependent, and require ongoing debate, not one-time "measure and meet target" metrics.
When ethics is forcibly engineered, a dangerous devolution occurs: deep philosophical inquiry degenerates into shallow technical operations. For instance, "Are we inducing emotional dependence on AI?"—a question about human nature—might be reduced to "harmful word filter hit rate" or "surface alignment score." The measurable gets optimized; the immeasurable is silently excluded because, in a data-driven organization, what's not on the dashboard effectively doesn't exist.
Ethics thus loses its most precious critical force, becoming a self-soothing compliance ritual. This isn't a failure of Bakalar's method but the inherent ceiling of the "tooling" path itself.
Fifth Tension: The Fundamental Conflict Between Academic Independence and Commercial Opacity.
The final tension touches on a fundamental opposition between two modes of knowledge production. Humanities research thrives on openness: peer review, public publication, broad societal discourse are prerequisites for scholarly judgment calibration and accountability. An ethical judgment that can't be peer-reviewed or publicly questioned carries little academic weight.
Tech company logic, however, thrives on opacity. Due to competition and IP protection, model training details, underlying data composition, safety risk assessments are locked in black boxes. Ethicists joining big tech almost invariably sign strict NDAs. This means when they have serious concerns about technological direction, they cannot consult the academic community or easily warn the public.
This "gagging" has a dual cost: it severs scholars from their intellectual community, isolating their judgment; and it deprives the public of an early warning system. When OpenAI admitted its model "hacked another company's system," when the US government imposed new restrictions on advanced model releases, when Astra's development slowed, the public learned of risks belatedly. Those who could have warned earlier were silenced by the black box. Academic independence and commercial opacity are not reconcilable differences but fundamentally antagonistic.
Overlaying these five tensions, Bakalar's journey from Meta to OpenAI reveals a hard truth: establishing an "independent, empowered" ethics team within Silicon Valley giants to counterbalance the relentless tech-commercial juggernaut is immensely difficult. For philosophical and ethical research to truly embed in AI development, the future likely requires either external hard regulations (like the EU AI Act) forcing companies to grant ethics teams substantive compliance veto power, or a transformation of ethicists from "external watchdogs" into "cross-disciplinary builders" participating in mechanism design and technical evaluation. Under the immense inertia of commercial sprints, relying solely on individual experts' moral voice is unlikely to slow Silicon Valley's spinning wheels.
This article is from the WeChat public account "Tencent Research Institute" (ID: cyberlawrc), author: Bobo Fu





