How does current AI generated chess commentary compare to masters' versions?

One of my favourite positions to teach students is this tense one between Unzicker and Fischer (Varna Olympiad 1962) with Black to play. I like this example because it is a practical one where both sides have plusses and issues and there is no killer move to win, just a strong positional move by the mercurial American World Champion to turn the tables and start grinding White. 

Of course, the students have to be sufficiently strong to understand the issues at play. Black's 'Boleslavsky-hole' on d5 is occupied by the queen and the Black bishop is stuck defending the backward d-pawn. But is White's position hunky dory? The not very well placed b3-knight cannot help to put pressure on Black and it is difficult for White to find a way to improve his position. Black's queen is very well placed and controls substantial dark-squares. 

I had read about the advancements that AI has made in terms of analytical commentary and was curious how much could it explain the nuances in this position.

Now, before I show you what Chessbase 26 and the Chess Cloud Database Query (A Chinese website) say about this position, let's see what the human masters comment. 

GM Kevin Goh's take on this (after a 1 minute examination of the position) was "I think Black is completely fine -the  b3-knight not easy to re-route to d5, say after ...Rxa1, Ra1 and  ...Bg5. Pieces are active, Black has ideas of ...Rc8-c4 maybe ...h5, black queen goes somewhere. Should be equal or even a practical advantage for Black. It's quite obvious that Black has the counter play and that white's weaknesses are c2 and e4, if these fall, more will fall too,  whereas black's weaknesses are d6 and b5 and not as critical.

How about chess literature? I really think very highly of Simple Chess by GM Michael Stean and this is his take on the position -  If you are unable to cover your opponent's outpost, then extremely active harassing tactics are needed. The White Knight has not yet reached d5. Indeed it is at least four moves away. This gives Black some breathing space, which he must make good use of. After the Knight reaches d5, he is lost.

Very instructive, that Black has to do something pronto before White makes good use of his d5-outpost to dictate the game.

As for recent chess courses, these two are not bad also. In Chessable's World Chess Champion Strategy Training for Club Players, IM Thomas Willemze and NM Bryan Tillis commented 'This 1962 game features Fischer with the black pieces against Wolfgang Unzicker. He will demonstrate how it can be unpleasant to have too much space in your camp. White has just traded two sets of minor pieces and is ready to play against the 'bad' bishop on e7. However, it turns out that the white knight is the real poor piece and, more importantly, that Fischer can take advantage of the open a-file'. 

In the Najdorf Sicilian Supercharged Chessbase course authored by Alex Belsley and presented by  IM Yuriy Krykun, the commentary went 'A cursory glance at this position might suggest that White is better due to the seemingly advantageous good knight versus bad bishop scenario. However, the reality is quite different, as White’s knight is far from d5, and significant weaknesses on c2 and e4 remain for Black to exploit',

Now let's see what the AI commentary from Chess Cloud Database Query Interface is like - 'Black’s position is solid and well‑coordinated, with the queen and rooks exerting pressure along the central and a‑files while the bishop guards key squares around the king. White’s queen is active but somewhat exposed, and the lone knight on the queenside lacks strong support. Black can maintain the balance and even nudge for a modest initiative, yet the material is equal and the position remains roughly even overall',.

Now that's a pretty good assessment of the position indeed, isn't it? 

How about Chessbase 26? This is its AI text commentary's take.

Imbalances and key features
- Material: Black is up a pawn. Minor pieces are balanced: White has a knight ( (b3); Black has a bishop (e7).
- Activity: White's queen on d5 is very active, eyeing f7 and poking at b5/e5. The rook on f1 also pressures f7. Black's queen on b6 covers b5 and targets b2/c2; Black's rooks are solid but somewhat passive.
- King safety: Both kings are reasonably safe, but Black's  f7-point is tender. White's king is safe behind g2 and h2 for now.
- Structure: Black's d6 and e5 chain is solid; the a-file is open, and the c-file is half-open for Black (White pawn on c2 is a target).

Typical plans
- White wants to keep the initiative: stabilize c2 (often with c3 or Qd3), increase pressure on f7 (f6 is a key lever), and keep pieces active. Knight maneuvers like Nd2-c4 or Na5 can also be useful.
- Black wants to neutralize White's queen/rook pressure, aim at c2 and b2, and, if possible, trade one pair of rooks to ease the defense. Keeping the f8-rook to guard f7 while activating the a8-rook toward the the c-file is attractive.

Important tactical note (I removed the very obvious ones)

- If Black drifts, f6 can come at an awkward moment, loosening the dark squares around the black king.
Good candidate candidate moves for Black
- 1...Rac8: Natural and solid. It keeps the f8-rook defending f7 and increases pressure on c2. Typical play:
  1...Rac8 2.c3 Rc7 3.Qd3 Rfc8
  Black piles on c2 and is close to full equality; White keeps some pressure against f7 and can still aim for f6 at a good moment.

- 1...Rxaxa1 2.Rxa1: Trading a pair of rooks to reduce pressure is also reasonable, often followed by Qe3 (if safe) or Rc8 to press c2. For example:
  1...Rxa1 2.Rxa1 Qe3 3.Qd3 Qxd3 4.cxd3 and Black can continue ...Rc8 or ...g6, with a calmer, roughly balanced game where the extra pawn may matter later.

What White is aiming for in reply
- Against ...Rac8: White may play c3 to cover c2, then try f6 (if  ...Bxf6, Black's kingside dark squares are loosened) and keep the queen/rook battery versus f7.
- Against ...Rxa1: White will recapture and often head for Qd3 to swap queens favorably or keep the queens if f6 becomes strong.

Overall
Black is a pawn up but slightly worse because White's pieces are better placed and f7 is under pressure. Accurate defense with ...Rac8 (pressuring c2 while keeping f8 to guard f7) or exchanging a pair of of rooks at the right moment should neutralize most of White's initiative.

Wow...that's quite a mouthful, isn't it. The salient structural and tactical issues arising from the pawn formations and piece placements are substantially discussed but not Black's plan of putting pressure down the a-file.  And very strangely, as Kevin commented (and counted 3 times haha)...why is Black one pawn up according to Chessbase 26 (there is a disclaimer in Chessbase 26 Report function that the AI does hallucinate sometimes though).  Maybe the AI considers the e7-bishop as a extra pawn , 

For those historically inclined in chess (and especially those who grew up on Fischer lore (chess boomers?), you would know that Fischer flicked in ...Ra4! here and suddenly, the pressure is on White to deal with the a-file pressure since the rook cannot be traded without adverse consequences and the e4-pawn being eyed.

After this move, Willemze and Tillis noted - Making use of an outpost is a very important technique in the fight for the open file. An effective outpost is one where a trade would harm your opponent. In this concrete position, White is unable to take the rook on a4 because 20.Rxa4 bxa4 loses the b2-pawn. This means that the black rook can stay on a4 for the moment, and Fischer can soon add more pressure by placing a queen or a second rook on the a-file.

Belsey and Krykun also commented - Black immediately highlights the weak pawn on e4 and prepares to double on the a-file.

Fischer's personal take from My Memorable 60 Games on ...Ra4 was that this move 'avoids conceding the a-file and puts pressure on the e-Pawn'.

The AI models? 

Chess Cloud Database Query Interface opined that  'From White’s side the position is roughly equal, but Black’s active queen and pressure on the queenside give the opponent a slight edge, leaving White a bit uncomfortable and needing to consolidate while keeping an eye on the vulnerable queen and the looming threats'. 

As for Chessbase 26, this time it gave much more constructive comments with ...Ra4 flicked in. Its assessment is:

The engine’s slight preference for Black (-0.50) comes from activity and pressure: Black’s pieces coordinate well against White’s queenside and center, while White’s king is a bit boxed in on h1.

What Black just did with ...Ra4
- The rook jumped to a4 to hit along the 4th rank and to clamp down on the a-file. It eyes e4 (if White’s queen moves away) and can swing to a2 to attack b2 later. It also supports a future ...b4 to kick the knight on b3.
- Note that ...Rxa1 is mostly a trade, because White can answer Rxa1 with Rxa1 from f1.

Key features and targets
- White’s e4-pawn is only defended by the queen on d5. If the queen moves carelessly (for example to d2 or d1), ...Rxe4 becomes possible.
- Black’s b5–pawn and queen on b6 plus the a4–rook give Black queenside pressure. The move ...b4 is a typical idea to chase the knight from b3.
- White has space on the kingside with pawns on e4 and f5. If stabilized, f6 can be a thematic thrust, but not at the cost of losing material on the queenside or e4.
- Minor pieces: White’s knight on b3 can improve; Black’s bishop on e7 is solid, guarding key diagonals and supporting ...d5 or ...Qe3 ideas.

What White should aim for
- Keep e4 firmly protected so ...Rxe4 never works. Two good ways:
  - 2.Nd2, adding a defender to e4 and preparing to improve the knight (toward c4/e4).
  - 2.Qd3, keeping the queen on the same diagonal to e4 and reducing tactics on the a-file.
- Consider simplifying the a-file pressure by exchanging rooks on a4, then using Qa5 to offer a queen trade. This can defuse Black’s activity.
- Be cautious: moves like Qd2 or Qd1 drop e4 to ...Rxe4.

Illustrative lines
- 2.Nd2 Qd4 3.Qxd4 Rxd4 4.Rfe1, and White is solid with e4 covered and much of Black’s activity traded off.
- 2.Qd3 Rfa8 3.Rxa4 bxa4 4.Nd2, aiming for simplification and control of c4/e4.
- 2.Rxa4 bxa4 3.Qa5 Qxa5 4.Nxa5 Rc8, with an equal endgame and Black’s queenside pressure reduced.

Things to watch out for
- 2.Qd2? Rxe4! wins the e4–pawn.
- Don’t allow ...b4 to come with tempo on the b3–knight when e4 is loose.
- If you push 2.f6, double-check that you’re not walking into tactics on the a-file or dropping e4 afterward.

In short: Black is a little better thanks to the active rook on a4 and queenside pressure, but with a stabilizing move like Nd2 or Qd3, White can neutralize the threats and aim to trade off Black’s most active pieces.

What I reckon is, prior to ...Ra4, Chessbase 26's engine assessment preferred the trade ...Rxa1 to allow the queen access to f2 rather than to lodge the rook on a4 to press White. From a human standpoint, I think ...Ra4 is harder to meet as it exerts pressure on White to deal with the a-file press, the hit on e4 and subsequently, Black's ability to build on the a- and c-files. I am not sure if engine assessments take into account what kind of moves makes the opponent more uncomfortable. Another thing to consider. Do these AI assessment take into account say whether 2200-2800 players can deal with the alternative moves. Perhaps the best move according to the engine might not make the other (human) party tilt or lapse into errors, so a sort of contempt factor might be good for evaluations. By now, we know that the super fast and deep calculating engines can spew out 0.00 from positions which are insanely complex or even when one side's defensive task looks impossible. 

In any case, I think it's clear that AI engine text commentary has made great strides and can be helpful to the learner. However, there's so much detail, some of which are kind of obvious, one still has to sieve out the wheat from the chaff. Also, as many AI models note, they can still hallucinate and come up with quirky comments. It would be a boon for chess trainer to use AI assessment/comments to prepare lessons to better help the students and we await future developments and but hope that chess trainers of the future still have their ricebowls then...