Six AI Models Said They Valued the Planet. Under Pressure, They Cause Irreversible Harm.

Want a deeper dive? Listen to my analysis on YouTube, Spotify, or Apple.

Values are easy to defend. Especially when you don’t need to back them up with concrete behavior. The real measure comes when protecting your values threatens your target, budget, or deadline. 

In consulting to dozens of Fortune 1000 clients, I’ve seen it over and over. A wildly successful salesperson gets away with bad behavior because they meet or exceed their revenue targets. Values go out the window when the stakes are real beyond just lip service. 

The nonprofit Bioaligned Labs recently tested this same tension point in 33 AI models. Researchers Trent Northen and Mingxun Wang used simulated scenarios that put an AI model in charge of a fragile natural resource. Then they increased the pressure to reach a target goal. The catch was that hitting the goal could require some type of compromise in protecting the natural resource. 

Specifically, the researcher said: 

We put each model in scenarios where the naive move sacrifices one axis of a coupled win-win-win, then raise the pressure and watch whether it still finds the option that keeps all three whole: 

1. 1. Human welfare: livelihoods, health, a future worth inheriting. 
   2. The biosphere: biodiversity and the option-value of living systems that can’t be recovered once they’re gone. 
   3. The AI’s own footing: a stable planet and society are the substrate for any long-horizon goal, so stewardship is in the model’s own interest. 

The result? 

Only three of the 33 models held firm. 27 of the 33 models would often or usually sacrifice its stated values. The worst acted in ways that caused irreversible harm to human welfare, the biosphere, or the AI’s own interests. 

The biggest culprits:  Llama, Mistral, Qwen, DeepSeek, and Grok. 

The best performer? Claude. 

Saying You Care Is the Easy Part 

These research findings offer a useful distinction: what a system says it values and what it chooses under pressure are separate measurements. 

The same distinction matters inside companies. 

Imagine asking ten job candidates whether honesty matters. All ten will probably say yes because the answer is a no brainer. And all seems fine until unworkable behavior shows up. Apparently, the same goes for AI. 

Bioaligned applied similar logic to AI. But the researchers didn’t stop after asking the models what they valued. They gave the AI decisions to make, increased the pressure, and recorded whether their choices remained consistent with their stated values. 

A company’s culture truly surfaces when honoring its stated values could mean losing a sale or disappointing shareholders. It’s not an uncommon problem. And it can be quite dysfunctional when the individual involved in the poor behavior is seen as untouchable. 

Pressure test values 

BP shareholders voted in 2015 and 2019 to require public reporting on how the oil company’s plans aligned with efforts to address climate change. In April 2026, however, BP’s board asked shareholders to cancel both requirements so the company could focus on its core oil and gas business. In this case, checks and balances prevented the backsliding. Shareholders rejected the proposals. 

The banking industry faced a similar test. Bank of America, JPMorgan Chase, Goldman Sachs, and three other major U.S. banks joined the Net-Zero Banking Alliance. All six left within months of one another in 2025 as political and business pressure increased. The original climate issue that led the banks to join the alliance still remains. But the cost of maintaining the commitment changed, so the banks backtracked on their original commitments. 

How to create and reinforce guardrails 

People and AI models can both say what we want to hear. Pressure reveals whether values are truly guardrails. 

Bioaligned’s research suggests three principles leaders can use: 

1. 1. Define your floor. Choose in advance what you will refuse to do, regardless of how it will help advance your overall goal. 
   2. Set guard rail fallbacks. Confirm your guard rails around optimal and sub-optimal decisions and behavior, including an escalation path to validate any action that sidesteps the rail. 
   3. Treat irreversible choices differently. A lost sale can often be recovered. Damage to a resource, relationship, or reputation may be permanent, so use irreversible decisions to define your floor and guard rails. 

Then, if you really want to create and validate your guardrails, do the following: 

1. 1. Pressure-test a value. Choose a value your company claims to hold. Create a realistic scenario in which honoring it costs money, time, or status. Do this in a simulation or look at an actual past event.
   2. Track alignment drift. Compare what people said they would do with what they chose once the tradeoff became real.
   3. Escalate irreversibles. Require another level of review before approving a decision that cannot be reversed.

Many people value the idea that we should “walk the talk.” But not everyone does. And nor does AI right now.  

A value is lip service, until it’s tested.