Conduct is not risk

Thailand’s prisons held about 301,020 people in 2025, in a system with capacity for 248,333.

That is roughly a fifth more people than the buildings were designed for, and it is the pressure behind everything else in this article. A system over capacity has to release people, continuously, and the question is not whether it releases them but how it chooses.

It chooses using a grade.

The six classes

Every sentenced prisoner in Thailand holds a classification — ชั้นนักโทษ — running from ชั้นเยี่ยม, excellent, down through ดีมาก, ดี, กลาง, and two grades below that.

The class is the currency of the entire system.

Remission is calculated from it directly: ชั้นเยี่ยม earns five days off each month, ชั้นดีมาก four, ชั้นดี three. Eligibility for parole depends on it. Eligibility for external work programmes depends on it — article 126 There is no employer described one, where a day worked outside was a day off the sentence. And where periodic clemency instruments apply, the scale of benefit is graded by class as well.

Your class is, in a real sense, your release date.

So what does the class measure?

What a class is awarded for

Conduct inside the prison. Compliance with regulations, cooperation with staff, participation in work and training, absence of disciplinary incidents, orderliness over time.

Every one of those is a reasonable thing for a prison to want, and every one of them is measurable. A prison can observe them, record them, and review them on a schedule. As an administrative instrument the classification does exactly what it was built to do, and it does it consistently.

The difficulty is what it is being used for.

A class is a record of how somebody behaves in a controlled environment with no alcohol, no money, no relationships of the kind that produce most violence, no opportunity, and a staff member within shouting distance.

A release decision is a prediction about how somebody will behave in an environment with all of those things.

Those are not the same question, and the classification only answers the first one.

Which is article 119, in a building

That piece looked at why Thai drivers form confident theories about other drivers from the make of car, and found the answer was not prejudice. It was that in a system with almost no visible driver record, the vehicle was the only variable available. People reason from a bad signal when there is no good one.

A prison has the same problem with better paperwork.

Risk is not observable from inside a cell. What is observable is conduct, and conduct is recorded meticulously, so conduct becomes the number that decisions attach to. Not because anybody believes it is the same as risk. Because it is what there is.

And a calm, cooperative, punctual prisoner who committed a violent offence in circumstances that will exist again on the outside scores well on the only instrument the system runs. Nothing in the classification is designed to notice that, because the classification was not built to notice it.

That is the gap, and it is a design gap rather than a scandal.

What is missing has a name

Comparable systems elsewhere run something the Thai system, on the published record, does not appear to: a structured risk assessment, conducted before release, distinct from the disciplinary record, asking specifically how likely this person is to reoffend and in what circumstances.

Those instruments are imperfect and contested. They encode their own biases, they misclassify people, and there is a serious literature arguing they should be used carefully or not at all.

But they are asking the question. A conduct grade is not asking it, and cannot be made to ask it by being calculated more carefully.

I want to be precise about the limits of what I am claiming. I have not found evidence that Thailand runs no risk assessment of any kind, and it is entirely possible that instruments exist within Corrections that are not published. What I can say is that the published criteria on which release benefits are calculated are conduct criteria, and that I could not find a published risk instrument feeding into them. That is a statement about the record, not about the institution’s intentions.

The number nobody agrees on

Reoffending is where this argument should be settled, and the Thai figures do not agree with each other.

Corrections statistics reported for 2016 give 62,117 reoffenders out of 261,687 sentenced prisoners — about 23.74%.

A Thammasat panel in July 2026, discussing overcrowding and reoffending, was reported as putting reoffending after release at 80%.

Those cannot both be describing the same thing. One is very probably a stock measure — the share of the current prison population with a prior conviction — and the other a flow measure over some follow-up period, or a different definition of reoffending entirely. I have not been able to reconcile them and I am not going to pick the one that suits the argument.

That, in itself, is the finding underneath the finding. A country cannot evaluate whether its release decisions are working if the reoffending rate is reported anywhere between a quarter and four fifths depending on who is speaking. The instrument is weak and the feedback loop that would reveal it is weaker.

The same July 2026 discussion put the economic cost of overcrowding and reoffending above ฿300 billion, which is the sort of number that gets a headline and which I would treat as an order-of-magnitude argument rather than an accounting figure.

Why the pressure makes it worse

A weak instrument under no pressure is a tolerable problem. A weak instrument carrying the weight of a system a fifth over capacity is a different thing.

When the buildings are full, the classification stops being one input among several and becomes the mechanism. It is the thing that can be applied at scale, consistently, to hundreds of thousands of files, without individual assessment. Its greatest administrative virtue — that it works mechanically — is exactly what makes it load-bearing when the system is under strain.

And overcrowding is not primarily a violence problem. Thai commentary is consistent that the population is heavily shaped by drug offences, which is a sentencing-policy question rather than a release-policy one. The people the classification struggles to assess are a small share of a very large number, and they are inside a machine built for the large number.

What would change it

Not longer sentences. This article is not making that argument and it should not be read as making it. A system that releases people badly is not improved by holding everybody longer; it is improved by choosing better.

A published risk instrument, separate from the disciplinary grade, applied to the offences where the question actually matters. That is a small number of cases relative to the population, and it is the number where the consequence of getting it wrong is highest.

And one reoffending series, defined once, published annually. Until that exists, nobody — including the department itself — can tell whether any change made anything better.

Common misconceptions

“Prisoners get out early because of leniency.” Release timing is calculated from an administrative classification with published rates: five days a month at the top grade, four and three below it.

“The class reflects how dangerous someone is.” It reflects conduct inside prison — compliance, work, absence of disciplinary incidents. It is not a risk assessment and was not designed as one.

“Thailand’s reoffending rate is known.” Reported figures range from about 24% to 80% depending on source and definition, and this article does not reconcile them.

“Overcrowding is caused by violent offenders.” Thai commentary consistently identifies drug offences as the dominant driver of the prison population.

“The answer is longer sentences.” This piece argues the instrument measures the wrong variable, not that the outcomes are too short.