an ordinary chessplayer wrote on 02/15/26 at 20:40:21:

aoc - I noticed that last night as well. Perhaps ChessBase thought the initial screenshot of the German-language AI report would be more accessible if it was in English instead. Whatever the reason, the English version shows differences to the original German. The evaluation is -0.18 in English, while it was +0.27 in the German AI report. Also 10...Qxc3+ is given more attention in the English version, with the AI thinking that 11.Bd2 would be adequately answered by 11...Qb2! (exclam apparently by the AI, which envisions that White will play 12.Qc1, exchanging queens while two pawns down). It's obvious that the AI isn't consulting an engine since it goes on so much about 10...Qxc3+, while a quick check with Stockfish shows the capture to give White a decisive advantage after 11.Bd2 Qb2 12.Rb1. Amusingly, the English AI report gives 10...0-0-0 as "a very logical alternative" to 10...Qc3+, seemingly totally oblivious to the fact that this move hangs the black queen.

I haven't run the "Consult AI" function more than once on the same position, so I haven't figured out if the differences between the English and German AI reports are simply random or if they're due to switching between languages.

aoc -Thanks for that Wayback Machine link and for your explanations of the workings of AI. 

Now that I can see the review's screenshot again, I'm disappointed that my memory exaggerated the difference between the Stockfish and AI evaluations. The Stockfish window says +-(1.66) with a comment that I translate to "White is clearly better". The ChessBase AI gives an evaulation of +0.27 while remarking, "So only a very small advantage, practically almost equal". That's still a big difference: the engine says White is close to winning, if not definitely so, while the AI has the position as nearly equal. However, as I understand the explanation, the AI doesn't actually consult an engine; it just happens to have one or more items in its knowledge base where a variation in the B20 Sicilian (and not necessarily the one under consideration) is evaluated as +0.27.

With aoc's explanation, the AI's confusion with the Winawer position from my initial post becomes more understandable. There are many Winawer Poison Pawn lines where the black rook ends up on g8, thus depriving Black of the privilege of short castling. The AI regurgitates this tidbit of information in the diagrammed position even though the black rook hasn't moved to g8. In fact, the AI seems to be incapable of realizing this.

Dink - I had to smile when you mentioned the Chessbase NFTs. They seem to have disappeared as soon as they were announced, though in checking my ChessBase account I see I have a "trading card" that I received as a reward for a purchase.

From what I have read about the ChessBase developers, they seem to be excited by new stuff and don't need management to hit them with a stick. They probably see the LLM as the new toy and are having fun with it.

Dink Heckler wrote on 02/13/26 at 21:50:27:

It might be the perfect fit in an obverse way. When the LLM hallucinates in another domain, say by quoting a non-existing precedent in a legal argument, it takes real effort to show the non-existence. But when the LLM hallucinates by miscounting pawns, the mistake should be instantly obvious to anyone who knows what a pawn is. Thus chess can be the public demo for just how inept these things can be with any technical subject.

And if you think about it, even in a non-technical subject what they say is merely plausible. Like the know-it-all at the dinner table latching on to whatever subject comes up and bloviating at great length, without regard for accuracy or insight. In that perverse sense computers have already caught up to the human brain.

cathexis wrote on 02/10/26 at 14:36:01:

The knowledge base is just a linking of tokens by proximity. So in the Winawer position Nernstian59 gave, it said black is a pawn ahead because in many Qg4 Winawer lines black is a pawn ahead. So usually the tokens "Qg4 Winawer" are near the tokens "black is a pawn ahead", and that's all it "knows". It has no independent insight into the Winawer and certainly cannot count pawns. And even though it says Stockfish is an engine, it doesn't know what that means either.

It's like a teenager whose brain isn't fully developed: 99 percent of the time they can behave responsibly, and 1 percent of the time they can do something that will get themselves or somebody else killed. But the AI is in a way worse, because we (well, other people) have been conditioned to believe the computer doesn't make mistakes, whereas with teenagers we know enough to give them a curfew.

cathexis - You're welcome! That's a good question about having the AI confirm its analysis by consulting an engine. This very point came up in a review of ChessBase 26 posted just yesterday on ChessBase's own website. The reviewer, Lukas Köpl (described as "a passionate club and tournament player"), had this to say about the AI in Chessbase:

"...one problem I encountered repeatedly was discrepancies between the AI's statements and ongoing engine analyses. For example, positions were rated only slightly better by the AI, whereas Stockfish clearly rated a position as +2. At times, the AI assistant's statements seemed generic and generalized. It would be interesting at this point to link the AI statements to the specific variations during an engine analysis in order to provide realistic interpretations of the ongoing analysis."

When I first saw the review yesterday, it had a screenshot, presumably from Köpl's computer, showing the AI report with Stockfish running in a separate window. The engine window showed a +3 evaluation, along with a comment that my high-school-level German translated as "White is clearly winning". The AI report had a statement to the effect that "White has a very slight advantage". These are obviously very different conclusions. The screenshot was gone today. I'm suspicious that it was removed because it showed the AI feature in an even less favorable light than the reviewer's text comments did. Also, I'm wondering if the screenshot showed some AI hallucinations that weren't brought up by the reviewer.

I recently tried the Consult AI feature in ChessBase 26. It's supposed to examine a position and provide a natural-language explanation of plans and strategic characteristics. The position I used arises after 1.e4 e6 2.d4 d5 3.Nc3 Bb4 4.e5 c5 5.a3 Bxc3+ 6.bxc3 Qc7 7.Qg4 f5 8.Qg3 cxd4 9.cxd4 Ne7 10.Ne2.



This position from the Classical Winawer is of interest since 10.Ne2 has scored extremely well (about 72% for White) and has put the line on life support for Black. I was curious if the ChessBase AI would have any new or useful ideas for the second player.

The Consult AI function took a few minutes to generate an output, and the result began with an overview of the position. The AI noted that Black was a pawn up. I did a double take and counted just to confirm that both sides had seven pawns. The AI then went on to say that the position had the characteristic French Winawer dark-square pawn chain on d5, e6, and f5. (Italics mine).

Next, the AI suggested that 10...Qxc2 should be checked as a possible plan, which seems reasonable enough, but then it added that Black can go up two pawns with ...Qxb2, forgetting that White's b2-pawn is longer there. 

Returning to the diagrammed position, the AI stated that short castling wasn't longer possible because the black rook was on g8. On the other hand, it said Black needn't worry about Qxg7 since ...Rg8 would trap the white queen, which isn't true.  Apparently the AI isn't quite clear where that black rook is. After deciding that short castling was impossible, the AI went on to outline a plan for long castling. (Note that "impossible" or not, 10...0-0 is by far the most common move in the position shown ).

The one sensible plan presented by the AI was ...b6 followed by ...Ba6 to exchange off the "French bishop". 

The heading for the "Consult AI" page in ChessBase 26 says, "AI can give valuable insights into the plans of an opening variation, but it can make mistakes. Check every statement carefully!" In this case, the warning is an understatement. It seemed like there were more hallucinations than useful information in this report.

In an earlier thread,  I linked to a video were ChessBase programmer Matthias Wüllenweber presented the results of a "Consult AI" report. Perhaps it was luck, or maybe it was a cherry-picked example, but the AI seemed much more reasonable there and no hallucinations occurred, although Wüllenweber did warn about that possibility at one point.

https://en.chessbase.com/post/matthias-wuellenweber-on-all-new-functions-for-che...