Same question, two answers
Ask an AI whether something is safe, and its answer depends on what it's allowed to look at. This story follows one carbon credit purchase to show why. No AI or climate background needed.
Elena, her company, and the Cedar Hollow project are fictional. The wildfire and buffer pool facts are real and sourced.
Elena has until year-end
Her company can't cut all of its emissions this year, so it has promised to make up the gap by buying carbon credits. Her annual climate report is due in December. The purchase has to happen before then.
A credit pays a forest to keep standing
The catch: if the forest burns, its carbon goes back into the air, and the credit no longer means what it says. To cover this, projects set aside spare credits in a shared reserve called the buffer pool, used to replace credits lost to fire.
Cedar Hollow ticks every box
Elena finds a forest project in Northern California. It passes the checks her company uses for every purchase.
- Certified by a major registryyes
- Price within budget$18 / t
- Enough credits available8,200
- Long-term commitment100 yrs
Before signing, she wants a second opinion.
She asks an AI assistant
What comes back depends on something Elena can't see from the chat box: whether the assistant answers from what it already knows, or from documents and data about this specific project. Let's look at both.
From memory: reassuring
Cedar Hollow is certified by a major carbon registry.
Forest projects are one of the most popular and trusted credit types.
If trees burn, an insurance reserve called the buffer pool replaces the lost credits.
A language model learns general patterns from its training. Asked about something specific, it describes how things usually are. Nothing here is false. It's generic, and it can't know what happened to this forest.
Grounded: the AI reads first
A better-designed assistant fetches evidence before it answers. Here's what happens behind the chat box.
- 1Elena asks the same questionNothing changes on her side.
- 2The product fetches sourcesIt searches for documents and data about this project.project documentfire trackerbuffer pool status
- 3The AI reads the question and the sources togetherIt's told to answer from what's in front of it.
- 4The answer comes back with citationsEach claim points to the source that supports it.
Retrieval is step 2: finding the right documents. Grounding is steps 3 and 4: answering from them and showing where each claim came from.
Watch the evidence change the answer
Cedar Hollow is certified by a major carbon registry.1
Forest projects are one of the most popular and trusted credit types.no source
If trees burn, an insurance reserve called the buffer pool replaces the lost credits.3
38% of the project area sits inside a 2024 wildfire perimeter.2
Tap a sentence to see its source.Commitment: 100 years
Location: Northern California
The 38% overlap is part of the fictional example. The buffer pool figure is real (CarbonPlan, 2025).
Grounding did four things
The AI didn't get smarter between the two answers. It got better information, and that changed each sentence differently.
Passing every check isn't the same as being safe
Elena's checklist asks whether the credit is legitimate. It never asks what could go wrong.
- Has the project burned recently?yes, 38%
- How fire-prone is the region?high
- Is the backup reserve healthy?~40% lost
A checklist asks yes or no questions. Risk analysis asks how likely a bad outcome is and how much it would hurt. A good grounded assistant surfaces both.
Offset forests are burning now
Cedar Hollow is invented. The pattern isn't. Forests sold as carbon credits in the western US have burned in several recent fire seasons.
It isn't new either. In 2021, forests that had sold credits to Microsoft and BP burned in Oregon and Washington. Microsoft's carbon program manager said at the time that the company had bought forest offsets that were now burning.
The risk was visible in public data. That doesn't mean every loss was avoidable. Fires are hard to predict project by project.
The real decisions are product decisions
The model was the same in both answers. What changed was the product around it. Three choices made that difference.
It's the same principle as the agent in piece one: show your work.
Same question. Different evidence. Opposite answer.
An AI's answer can only be as good as what it's allowed to look at. Deciding what it looks at, and how it shows you, is part of designing the product.