Online Chess After the Pandemic: When ACPL Cannot Judge a Human Being

Game six of the 2026 World Chess Championship in Dubai lasted 136 moves, close to eight hours. I sat in front of a screen in Saigon from 7 p.m., writing every move into a squared notebook, and by the time only kings, rooks and a few pawns remained, dawn had broken. Magnus Carlsen drew with Ian Nepomniachtchi in the longest game in the history of world championship matches.

What I remember is not the number 136. It was a position with four rooks and six pawns, where every engine evaluation sat uncomfortably close to the balance line, and both players made almost no error large enough to name. When the game ended, I opened the forums. Nobody discussed move 89 or move 112. They discussed why nobody had blundered.

That was the moment I understood how audiences read a chess game had changed. After every game in any online event, the first question now concerns integrity. And it is being answered with a metric originally designed to measure the quality of moves.

The board leaves the playing hall

In March 2026, the entire global chess calendar stopped within weeks. No spectators, no metal detectors, no arbiters walking between tables. FIDE had to stage an online Chess Olympiad for the first time in its near-century of existence.

What happened in the final of the 2026 FIDE Online Olympiad was the first lesson and also the most complete one. India met Russia. While the two teams traded points, two Indian players lost their internet connection and were defaulted under the published rules. India appealed. The organisers, after review, declared both teams joint champions.

That was a reasonable administrative decision under the circumstances, but it exposed a gap far larger than one final. The laws of chess were written for a physical environment where bandwidth does not exist as a variable. Once the board leaves the hall, new variables appear: latency, throughput, where the server sits, and most importantly, who holds the game log.

The money followed the same migration. Online events with weekly and monthly prizes attracted hundreds of grandmasters and thousands of rated players. Major platforms expanded prize funds, sponsorship and the betting markets that ride on the result of a three-minute blitz game. Within roughly two years, a significant volume of money flowed into a system whose oversight infrastructure barely changed.

In Vietnam, that wave arrived about half a year later but with no less force. Youth events moved online, clubs opened classes through screens, the number of competitive accounts surged. A new playing field formed in a very short time. But the number of arbiters trained in online anti-cheating grew far more slowly, and remains the structural weak point of the whole region.

I spent two years as a chess commentator for VTC starting in 2026, and I remember the feeling of sitting in a studio with a real board in front of me. Everything I needed was within sight. Ten years later, during an online broadcast at two in the morning, what I needed sat inside a browser window I had no right to access.

What ACPL measures and what it does not

ACPL, average centipawn loss, is the metric most often used when someone wants to argue a game looks abnormal. The calculation is not complicated. For every move, the engine gives its best evaluation of the position. After the player moves, the engine evaluates the new position. The difference between the two values, expressed in centipawns, one hundredth of a pawn, is the loss on that move. Averaged across the game, that is ACPL.

Its strengths are obvious: objective, reproducible, comparable across games and players. Its weaknesses are no less obvious, and are usually ignored in online arguments.

First, ACPL depends on search depth. The same game analysed at depth 18 and at depth 30 produces two different numbers, sometimes far enough apart to change the conclusion of a careless observer. Second, it depends on the type of position. In a locked position where most moves are nearly equivalent, ACPL is naturally low for both sides. In an open position with many options, ACPL is naturally high. Comparing the ACPL of a defensive game with that of an attacking game is comparing two different things.

Third, and this is the point I want to stress most: ACPL measures the quality of a move, not its origin. A good move made because a player understood the position and a good move made because an engine stood behind them carry the same centipawn value. The metric cannot distinguish between the two, and mathematically it was never designed to.

In other words, ACPL is a ruler for quality that has been dragged in to serve as a ruler for ethics. That is the fundamental methodological error in almost every online discussion of chess cheating I have read over four years.

The opening trap and engine match rate

The second commonly used metric is engine match rate, the share of moves matching the engine's first choice. It sounds more serious than ACPL, but in practice it is easier to contaminate.

Picture a grandmaster rated 2600 entering an opening line he has played three hundred times. In the first fifteen moves he does not need to think. He plays from memory. The match rate in that phase can reach one hundred percent without any assistance. So a threshold such as above ninety percent is suspect, when calculated across a whole game, is close to meaningless.

The right approach is to split the phases. Count only from move 20 onward, once the position has left recorded theory. Even then, the threshold must be relative. The correct question is not what percentage this player scored, but what percentage he scored compared with his own long-run distribution in the same type of position and the same time control.

I once compared two games by the same player, two months apart, same opening, same type of middlegame. The first game showed a 71 percent match rate, the second 94 percent. Look only at the second and the conclusion is very different. Look at both and it becomes clear that in the first game he left theory, and in the second he returned to his familiar rails. Nothing suspicious in either.

A heat map can lie, but five consecutive failed presses cannot. In chess, the equivalent sentence is: a beautiful ACPL figure can lie, but thirty consecutive games across different position types rarely can.

The sample size problem

This is where most online conclusions collapse. The probability that a 2500-rated player produces one extremely accurate game in a favourable position is not small at all. A single game is a single observation. To say anything statistically meaningful, you need a distribution, and a distribution needs time.

More precisely: you need at least thirty to fifty games by the same player, in the same period, at the same time control, before you can begin to speak of an anomalous pattern. Even then you must rule out obvious alternative explanations: increased training volume, a newly prepared opening, a peak form period, or simply a weaker opponent making the position easier to read.

Meanwhile, social media needs about twenty minutes after the final move to convict a player. This is the most dangerous asymmetry in the whole story: scientific procedure requires hundreds of games, while public judgement requires one game and one screenshot.

200 milliseconds and the limits of judgement

There is another variable that online analyses almost always ignore: time control. In blitz and bullet, a significant share of moves are executed by pattern recognition rather than calculation. The player sees the position, recognises the shape, and moves pieces almost reflexively.

Esports taught me something football hides well: a reaction within 200 milliseconds can break an entire system. In bullet chess, that means a technically flawless sequence can appear without any conscious thinking behind it. That makes the data look more suspicious while actually containing less judgemental information.

Here is the paradox: the games with the cleanest numbers are often the ones with the fewest cognitive traces to analyse. In classical chess, with two hours for forty moves, every decision leaves a trace in thinking time, in gaze direction, in how long a player hesitated before a critical move. In three-minute blitz, those traces nearly vanish.

So when someone sends me an analysis claiming a blitz game proves cheating, my first question is always about time control. If it is three minutes, I almost always set that analysis aside.

The Niemann and Carlsen affair: suspicion without a court

In September 2026, at the Sinquefield Cup in St. Louis, nineteen-year-old Hans Niemann beat Magnus Carlsen with the black pieces. Carlsen withdrew from the tournament. Days later, at another event, Carlsen resigned after one move when paired against Niemann. That same month, Niemann filed suit against Carlsen, Chess.com and Hikaru Nakamura in a federal court in Missouri, seeking one hundred million US dollars in damages. In June 2026, the judge dismissed the case.

The affair was discussed endlessly, but usually as rumour. The angle I care about sits elsewhere: for nearly a year, no independent adjudicating body published a verifiable conclusion on whether Niemann had received assistance.

A private platform produced its own report on closed accounts, but a document from a commercial company operating a product cannot substitute for a ruling. FIDE has a Fair Play Commission, but its jurisdiction is explicitly limited for events outside the FIDE system.

The result of that gap is a paradox: a young player was judged by public opinion, a world champion was questioned over his motives, and neither was exonerated nor convicted.

As someone who reads games, I consider this a greater loss than any specific cheating case. A cheating case detected and processed correctly makes a sport stronger. A suspicion with no mechanism for closure weakens it from within, because it turns every young player into a potential suspect and every spectator into a judge without jurisdiction.

Self-appointed statistical referees

After the Niemann affair, a new phenomenon appeared in chess: self-appointed statistical referees. Several prominent players, among them former world champion Vladimir Kramnik, began publicly suspecting a string of young players based on their own statistical analyses.

The phenomenon has two sides and I do not want to oversimplify. On the positive side, it forced platforms to become more transparent about their processes and calculations. Before that, there was almost no incentive to publish methodology.

The negative side is far heavier. When a famous player posts a chart on social media, he operates a parallel justice system with no appeal, no confidentiality, no investigative deadline and no jurisdiction. In a sport where personal reputation is the only asset, a public accusation can destroy a career faster than any formal sanction.

And there is a technical problem usually ignored in these debates: if a player is good enough to use engine assistance without detection at move level, then a statistical analysis performed by another player on public data is almost incapable of proving otherwise. We are using weak instruments to announce strong conclusions.

Data limits

Here I must be explicit, because I have kept this section in every analysis since 2026.

Every ACPL threshold or match rate I mentioned above comes with conditions. They depend on engine version, search depth, the number of moves included in the average, and the game phase chosen. An ACPL computed at depth 30 in 2026 differs substantially from the same game computed at depth 18 in 2026. Schedule conditions, rest gaps between rounds, and whether an event took place inside or outside lockdown all shift the standard distribution.

In 2026 I discarded half of my older data set, because chess after the lockdown is a different sport. Not because the rules changed, but because playing conditions changed. A player competing in three online events a week has an entirely different training rhythm and reflex rhythm from one competing in four over-the-board events a year. Comparing them means mixing two data sets generated under different conditions.

Saying this is not a rejection of statistical analysis. It is placing it correctly: a probabilistic instrument, not a court of law.

The view from the less-watched side

There is a group that debates about online chess cheating almost never mention: women players.

In women's events, prize funds are smaller, cameras fewer, arbiters fewer, and verification procedures are often applied more lightly. That means a thinner protective layer in both directions: protection against cheating, and protection against false accusation. A woman player suspected online usually has no communications team, no lawyer, no representative to respond within twenty-four hours.

I have followed several online women's events featuring Vietnamese players, including Phạm Lê Thảo Nguyên. The technical structure of these events is usually simpler, the organising team thinner, and much of the oversight rests on the goodwill of the parties. That is an asymmetry nobody created deliberately, but it exists.

And there is a larger blind spot covering everything. We are pouring resources into optimising detection tools while the root problem lies in the structure of incentives. Where there is betting money, there is pressure. Where there is no adjudicating body independent of the organiser, with investigative power and the right to subpoena data, suspicion automatically replaces judgement. Better detection tools only mean more suspicions are generated, not necessarily that more cases are resolved.

The transfer market is chess, not a card game, but plenty of sporting directors prefer to flip cards. In online chess, the equivalent is: cheating detection is chess, but many people are playing a card game by throwing a single metric at the public.

One more point. Over-the-board chess is not cleaner than online chess. It simply has better traces. Over-the-board events have cameras, arbiters, security checks, and a broadcast delay of several minutes for cross-checking. In online events, the traces live in server logs, which only the platform can see, and which are usually published only when the platform decides to publish them.

The difference lies not in the ethics of the players. It lies in the quality of the chain of evidence.

What to watch in the coming years

Three things will, I believe, decide whether online chess keeps public trust or loses it.

An appeal process with clear deadlines. Today a player under suspicion in an online event often does not know how many days they have to respond, who receives their data, and who will read it. Standardising this matters more than any detection algorithm, because it determines the speed and transparency of any ruling.

An adjudicating body independent of the organiser. The conflict of interest is stark: the platform organises the event, earns from it, investigates, adjudicates, and holds the data. No professional sport operates sustainably with that four-in-one structure.

Publication of anonymised data for community verification. If ACPL distributions from hundreds of thousands of games were published anonymously, everyone would share a common baseline for comparison. A suspicious chart would then sit beside a public standard instead of beside a feeling.

Without those three, every good game will become a trial without a judge, and every young player will carry a burden unrelated to chess: proving they are clean.

No tactic is ever old, only the way we read a game expires. The 136-move game in Dubai remains a masterpiece of endurance and technique, and it still stands after every argument about metrics has settled. What I carry after four years of reading debates about online chess cheating is a question without an answer: are we protecting the result of the game, or the right to believe in it?