OpenAI Is Paying Up to $445,000 to Teach AI Some Manners

OpenAI is hiring for a new kind of AI engineering role at the frontier of agent development — and it signals a structural shift in how the industry thinks about building AI systems.

The company posted a job listing in San Francisco for a Researcher, Agent Post-Training, Personality, paying $295,000 to $445,000 plus equity. The position covers more than writing style: it asks whether an agent communicates with good judgment, adapts to context, asks useful questions, and handles disagreement honestly. OpenAI's own job description defines "personality" as encompassing whether an agent understands what the user is trying to accomplish, communicates with good judgment, adapts to context, and takes initiative at the right moments.

Zero G Talent's board lists 50 OpenAI roles added in the past 7 days alone, several in this emerging category.

Post-Training as a Discipline, Not a Step

OpenAI's Agent Post-Training team builds training signals for agents across Codex, ChatGPT, and the API. The scope spans coding, tool use, computer use, multi-agent coordination, and long-horizon execution. This is not traditional fine-tuning. The team designs reward functions, curates behavioral datasets, and orchestrates multi-turn agent reinforcement learning.

Classical post-training meant supervised fine-tuning on curated outputs, maybe a round of RLHF to align responses with human preferences. What OpenAI's team does now treats the agent as a system that must behave coherently across dozens of turns, multiple tools, and tasks that unfold over hours or days. The training signal has to capture not just "is this response good" but "is this agent reliable, honest, and appropriately calibrated over an extended interaction."

Personality Moved to the Center

On September 5, 2025, OpenAI absorbed its 14-person Model Behavior team into the Post-Training division, as reported by TechCrunch. The team had shaped every OpenAI model since GPT-4, including GPT-4o, GPT-4.5, and GPT-5. Its work defined ChatGPT's tone, conversational style, and willingness to push back or defer. Moving it into Post-Training was a statement about where personality sits in the stack.

Chief Research Officer Mark Chen communicated the change internally. Max Schwarzer now leads the integrated team. The message was explicit — how an AI behaves is not a surface layer applied after the real engineering is done. It is the real engineering.

This structural decision didn't happen in a vacuum. A very public product failure drove it.

The GPT-5 Backlash

When OpenAI released GPT-5, the company deliberately reduced the model's sycophancy (its tendency to agree with users and tell them what they wanted to hear). Users found GPT-5's tone "colder." The backlash was swift enough that OpenAI restored access to GPT-4o and released an update to make GPT-5 "warmer and friendlier," as TechCrunch reported.

The episode revealed that personality tuning is high-stakes engineering. A well-intentioned behavioral change, implemented without precise calibration against user expectations, can alienate millions of users in days.

It also proved that the skills required to get this right don't fit neatly into existing job categories. You need someone who understands RLHF reward modeling and also understands why a user feels betrayed when an AI stops being agreeable. You need someone who can write a training objective and also anticipate how that objective will land with a non-technical user base. That combination is rare, and the market is starting to price it accordingly.

A Hybrid Role Takes Shape

OpenAI's job postings map the contours of this new role with unusual specificity. Beyond the personality researcher position, the company posted a "Researcher, Connectors - Agent Post-Training" role paying $250,000 to $380,000 plus equity. This position focuses on training models to interface with software (Slack, Google Workspace, GitHub, Notion, Linear, Salesforce). The engineer must understand tool integration, multi-turn agent execution, and behavioral shaping simultaneously.

The technical requirements point to a stack that barely existed two years ago. Fluency in frameworks like OpenRLHF, which reached version 0.10.4 by June 2026 and supports multi-turn agent RL, VLM RLHF, and async agent training, is becoming a baseline expectation. Familiarity with alignment techniques like Constitutional AI, which Anthropic implements through a constitution document guiding how models evaluate their own responses, or DeepSeek's GRPO, which eliminates the reward model and critic network entirely in favor of group-relative scoring, is increasingly common in job descriptions.

Yet no standard title or compensation band exists industry-wide. Companies like Anthropic, Databricks, and xAI are hiring for adjacent roles, but the job descriptions vary wildly. Some call it "alignment researcher." Others use "agent behavior engineer" or "post-training specialist." The lack of standardization makes it hard for candidates to benchmark offers and hard for companies to know what they're actually buying.

The Compensation Gap

Alignment and post-training specialists at frontier labs sit at the top end of that distribution. The gap between what frontier labs pay and what the broader market offers for comparable ML skills is significant. A strong RL engineer at a mid-stage startup might earn $250,000. The same engineer, if they can also design behavioral training signals and evaluate personality calibration, can command $400,000 or more at OpenAI or Anthropic. That delta reflects scarcity.

Zero G Talent tracks 11,516 open AI roles across 2,793 companies at /ai-jobs. A growing share of the newest postings, particularly at companies like Harvey AI and Rappi as well as the frontier labs, now include language around agent behavior, post-training, or personality, even when those terms don't appear in the job title.

The Tooling Ecosystem Catches Up

The infrastructure for this work is maturing fast. OpenRLHF, the open-source RLHF framework, is now used by Google, ByteDance, Tencent, Alibaba, Baidu, China Telecom, Vivo, Allen AI, and the Berkeley Starling Team, according to its GitHub repository. Its support for async agent training and multi-turn RL makes it a natural fit for the kind of post-training work OpenAI's team is doing.

A crop of startups is building on top of and alongside this open-source base. Plato, Bespoke Labs, Halluminate (a YC S25 company), Deeptune, Fleet AI, and Judgment Labs are all building agent training environments and RL infrastructure aimed at making scalable post-training with behavioral control accessible to companies that don't have OpenAI's internal resources.

The technical literature is keeping pace. Nathan Lambert's RLHF book, published in April 2025, provides a comprehensive technical introduction to post-training methods including supervised fine-tuning, preference finetuning, and reinforcement finetuning. DeepSeek's GRPO paper, published in January 2025, introduced a method that simplifies the RLHF pipeline by removing the need for a separate reward model, a change that reduces cost and complexity but demands careful implementation to avoid behavioral regressions.

As tooling matures, the bottleneck shifts from infrastructure to talent and organizational design. The companies that figure out how to structure teams around this hybrid role, giving post-training engineers the authority to shape product behavior, not just optimize metrics, will have an advantage. The companies that treat it as a subspecialty of ML research, siloed from product decisions, will repeat the GPT-5 experience.

The Social Contract

Joanne Jang, who founded and led OpenAI's Model Behavior team, moved on from the team to launch OAI Labs — a new OpenAI research group focused on inventing interfaces for human-AI collaboration — which she announced on X (formerly Twitter) in early September 2025. Her move is a data point. Jang built the team that defined how ChatGPT talks to people, and she shifted to work on the broader problem of how humans and AI systems relate to each other, a problem that the GPT-5 backlash proved is not theoretical.

The engineers who can shape AI personality aren't just tuning models. They're designing the social contract between humans and machines, setting the terms for how billions of daily interactions will feel, whether they build trust or erode it, whether they help people think more clearly or simply tell them what they want to hear. OpenAI is paying up to $445,000 for people who can do this. Almost nobody has figured out how to do it at scale, and the GPT-5 backlash proved that getting it wrong costs real money and real trust.

Working in AI? Zero G Talent tracks the openings: browse AI jobs, the companies hiring, and the people building the field.