Forest Carbon Credits Are Sold On a Counterfactual. A Counterfactual Is a Model.

On 9 October 2025, Science published an evaluation by Yuzhi Tang and colleagues of 52 voluntary REDD+ projects across 12 tropical countries.

Read the title carefully, because it is doing something most coverage of this subject refuses to do: "Tropical forest carbon offsets deliver partial gains amid persistent over-crediting."

Both halves. Partial gains. Persistent over-crediting.

The findings, precisely

From the Science study:

• Only a minority of project units showed statistically significant reductions in deforestation.
• Just 19% met their reported emissions targets.
• An estimated 13.2% of tradable credits were supported by counterfactual analysis.

From a Nature Communications study combining six independent evaluations of 44 REDD+ projects:

• The projects claimed, in aggregate, 10.7 times more avoided deforestation than independent estimates justified.

From a meta-analysis covering nearly one billion tonnes of credits — around 20% of all volume ever issued:

• Fewer than 16% of credits represent real emissions reductions.
• Forest conservation projects are the worst offenders, with overestimation ratios reaching 1:13.

And the 2023 investigation by The Guardian, Zeit Online and SourceMaterial found that more than 90% of Verra's rainforest offset credits likely did not represent genuine carbon reductions.

The nuance that both sides drop

The "90% worthless" headline travelled further than anything else in this story. It is close to what one investigation found. It is not what the peer- reviewed literature says, and the difference matters in both directions.

Against the offset industry: the over-crediting is real, large, systematic, and confirmed by multiple independent teams using different methods. This is not a smear campaign. It is a convergent scientific finding.

Against the "it's all a scam" framing: partial gains means gains. Some projects reduced deforestation, verifiably. A credit that overstates its impact by 10x is not a credit worth zero — it is a credit worth roughly a tenth of what was sold. Those are very different claims, and "worthless" is the wrong one.

The real finding is more damning than fraud and harder to fix: the accounting is broken in a specific, identifiable way.

Where the error actually comes from

The over-crediting is driven by selection bias in projects' control areas and modelling approaches.

Here is the mechanism, and it is the whole article.

A forest carbon credit is not a claim that carbon was stored. It is a claim that carbon would have been released and wasn't. The product is avoided deforestation — the difference between what happened and what would have happened otherwise.

That "otherwise" does not exist. It cannot be observed, measured, or audited. It has to be constructed — usually by picking a comparison area and asserting the project area would have followed its path.

Which gives whoever picks the comparison area an enormous, entirely legal lever. Choose a control that was about to be cleared, and your untouched forest looks heroic. You need not lie about a single hectare. Every observed number can be accurate and the credit can still be inflated tenfold, because the inflation lives in the counterfactual, not the measurement.

This is the same distinction we drew in Indigenous lands and Amazon deforestation: the observed 1.2% loss over 35 years is a fact; the modelled 35% counterfactual is an argument. There, the model was used to make a policy case, and we said to read it as a model.

Here, the model is the product. It is priced, sold, retired, and reported against a corporate net-zero target. A modelling assumption becomes an asset on a balance sheet.

A PNAS analysis makes the point at national scale, finding potential for strategic behaviour in jurisdictional REDD+ — the same lever, larger.

What is being done

Verra's revised standards, implemented in 2024, retroactively reduced the credited value of projects already in use.

Read that twice. Credits already issued — many already sold, many already retired against corporate climate claims — were revalued downward after the fact. That is a standards body admitting its earlier output overstated itself, and it is genuinely more honest than most institutions manage.

It also means a company that bought, retired, and reported those credits made a climate claim that has since been marked down. There is no obvious mechanism for unwinding a public statement.

The research literature is converging on a fix: systematic ex-post evaluation — judging projects on what actually happened afterward, rather than certifying a forecast in advance. Sell the credit after the forest is still standing, not before.

One commitment, since this is our field

We are building a platform in the same general space, so one line is worth putting on the record early — not as a pitch, but as a constraint we are accepting before we have anything to sell.

AEDO1 will not sell carbon credits, offsets, or any claim about avoided emissions. Not now, not later.

The reason is the whole article. An offset's value depends on a world that never happened. We would rather deal only in things that can be checked against something that did.

That is a much smaller commitment than it sounds. It rules out a revenue line; it does not earn us any credit. We are stating it now because a promise made before there is money on the table is worth marginally more than one made after.

Why this is on AEDO1

The voluntary carbon market is the largest existing attempt to make nature protection pay for itself. That is roughly the same problem we are trying to solve, and it is the closest thing to a cautionary tale we have.

Its central lesson is not that people are dishonest. It is that when the product is a counterfactual, the incentive to pick a flattering one is structural — and no amount of good faith survives an accounting method that rewards the assumption rather than the outcome.

If you build a conservation funding mechanism, build one where the number can be checked against something that happened.