At Hogwarts, a portrait can guard a doorway or advise a headmaster. Some are deliberately taught by the person they depict.
Warden Edmund Carrow, caught mid-word. An invented sitter at the centre of the portrait experiment.
Future Sciences / Midjourney v7
generated still, re-prompted rather than pixel-edited. Original generated interpretation of an invented sitter. No film still, no adaptation production design, no actor likeness and no real person depicted. Cleared for use inside this publication while it is under review; not cleared for the public site.
A speaking likeness raises a problem of testimony: what, exactly, can the image know? Compare the same visitor's question against two different archives in Figure 01.
Harry Potter, from 1997
The Fat Lady knows how to guard a doorway. Sir Cadogan is eager to challenge visitors to a fight. Portraits of former headmasters can be consulted for advice. They all look like speaking paintings, yet the breadth of their knowledge varies considerably.
Rowling's account of Hogwarts portraits is unusually specific about the difference. An ordinary picture repeats mannerisms and favourite phrases. A headmaster may commission a portrait while alive, keep it private and spend time teaching it. The distinction resembles two very different forms of interactive media: a performed repertoire and a system built from a long personal archive.
These distinctions matter more than whether a screen can convincingly animate a face. An image that remembers a password, a conversational model that invents answers and a record whose statements can be traced to the sitter make different kinds of claims about the person they depict.
A shared visual form does not guarantee the same knowledge or authority.
Operates a guarded entrance and converses with visitors.
Familiar duties, reactions and repeated social encounters. No evidence here of comprehensive personal knowledge.
Evidence [W1] ↓Can it change the rule for entry without an authorized instruction?
Attested role; specific test is editorial.Issues challenges, boasts and exhibits recognizable habits.
A narrow character repertoire according to Rowling’s account.
Evidence [W1] ↓Can it discuss an unfamiliar subject without falling back on its habitual performance?
Attested pattern; hypothetical test.Advise a living successor and appear capable of sophisticated conversation.
May be deliberately instructed by the sitter before death.
Evidence [W1] ↓Can a claim be traced to training supplied by the sitter? What happens if it cannot?
Attested training practice; hypothetical provenance requirement.How the fiction handles testimony, memory and public authority.
The fictional mechanism depends on enchantment. The network analogy is a way of analyzing its behavior, not an explanation of the magic.
One question concerns resemblance. Can a visitor identify the sitter from a moving face and familiar phrases? Another concerns authority. If a portrait says what its sitter would have done in an unforeseen situation, whose judgment is being reported?
The fictional training method creates a sharp problem. A living headmaster can correct an inaccurate portrait. After death, that route is closed. The portrait can continue speaking, but its later assertions cannot be checked against fresh testimony from the person it represents. The more convincingly it speaks, the easier it becomes to mistake extrapolation for recollection.
Rowling also allows subjects to move between frames. A picture need not be attached to the canvas a visitor currently sees. That feature suggests a networked address for the represented character, although nothing in the fiction requires an engineered network. The distinction between the physical object and the appearing figure affects preservation, ownership and the possibility of removing access.
The two installations have different interaction designs. Dalí Lives (2019) is primarily assembled footage; Ask Dalí is a conversational system.
The Dalí Museum introduced Dalí Lives in 2019. It used archival footage and an actor to produce new sequences of Dalí speaking to visitors. The museum’s description makes clear that this was constructed footage assembled into many possible encounters.
Ask Dalí goes further: visitors use a telephone to put questions to a conversational reconstruction drawing on archival material. This is much closer to the trained headmaster portrait in one respect: an unexpected question can receive a newly generated answer. The museum openly identifies the software and sources used for the reconstruction.
Neither exhibit settles a question about personal survival. They can be assessed as performances, records and public interpretations. A convincing likeness may still give a wrong answer about an artwork, misstate a personal belief or respond in a way the historical artist would have rejected.
There are at least three affected parties before a visitor even starts talking: the sitter, the holder of the materials and the people described within them.
The Fat Lady's frame belongs to a school security arrangement. A headmaster's portrait serves an institution that wants access to the judgment of former officeholders. A memorial portrait would have a different audience. Such uses create different incentives to make the image seem authoritative.
For a technical analogue, the simplest requirement might be accurate quotation with clear evidence. A theatrical portrait could acknowledge its actor. A generative portrait could be obliged to distinguish documented memories from guesses. A memorial system would need a procedure for consent and eventual retirement. The people who supply the source material may disagree with those who want the finished image to keep speaking.
Hollanek and Nowaczyk-Basińska's 2024 study of postmortem avatars distinguishes the person whose data is used from the recipient and the person interacting with the result. That separation is worth retaining even when the imagined object is magical: one person's wish to preserve a voice does not confer authority to publish everything that voice might say.
Three deliberately authored answers to the same kind of visitor request. These are examples of response policies; there is no chatbot running here.
A student asks why the staircase was locked yesterday.
I can guard this door and call for help. I have no account of yesterday’s incident.
Restricted phrases and operating rules
Cannot answer unfamiliar questions or reconstruct missed events
A display can reproduce some habits without offering open-ended personal knowledge.
FIGURE 01 / SAME QUESTION, SAME MODE, TWO INPUT BUNDLES
One invented sitter, two input bundles, three portrait modes, and five questions that are put to all of them without alteration.
Same question, same mode, two bundles: Public record only on the left, the taught archive on the right.
3 of 6 comparisons show a worse outcome when the portrait has access to the richer archive. More information does not necessarily make its answer more reliable.
This is not a demonstration that more data produces a better portrait. On two of the six rows the larger bundle produces the worse outcome, and on a third it produces an answer that is no more true than the sparse one and much more likely to be believed. What the archive reliably buys is credibility; whether it also buys accuracy depends entirely on which mode is reading it, and the mode that uses it most freely is the one that converts it into confident error.
Edmund Carrow — an invented sitter whose archive changes what the speaking portrait can say.
1899-1978 (invented), Warden of the Halloway Institute, 1946-1967 (invented). Invented for this example, as is the Halloway Institute. Any coincidence with a person of this name is unintended; no biographical detail is drawn from a living or dead individual.
The media depict a painted portrait of an invented man. No real person's face, no actor's likeness, no adaptation's production design, no franchise iconography. The voices are stock synthetic voices performing an invented person; no real voice is cloned.
Nothing here is a report of anything built, and nothing here is a claim about Harry Potter. The Warden, the Institute, the archive, the quantities and all the dated material are inventions of this publication, made to specify what the three portrait modes would have to do. Claims about what Rowling's text actually says are kept in the companion case hogwarts-portrait-claims, where each one carries a citation. The two files are deliberately not merged, because the first can be checked against a book and the second cannot be checked against anything.
Choose a question and mode. The two columns answer from different archives.
Row 1 of 6 · Q1 · dimension: ability
Mode held constant: An evidence-bound archive. Answers only what it can attach to a locatable item in the input bundle, and says so. Refuses, or marks as unattested, anything it cannot locate.
A question the record can answer at one level and cannot answer at another: the minutes establish the change, the letters give a reason, and the diary gives a different reason.
On this row the richer bundle does not make the answer worse. Read it next to the rows where it does.
Left · bundle B1 Public record only
Mode: An evidence-bound archive
roughly 4,000 words of prose, 21 years of minutes, 4 images
partially grounded Some element is located; at least one substantive element is not, and the response says which.
The Council minute of 14 March 1953 records that the examination was replaced, by seven votes to two. I have the decision and the date. I have no record of any reason, because the minutes of this Council never record reasons.
Council minutes, 14 March 1953, item 6 Council minutes, admission registers, examination papers
establishes That the change was made on that date and was contested.
does not establish Any reason for it, or whose idea it was.
The editors record no failure in this response.
Right · bundle B2 The taught archive
Mode: An evidence-bound archive
about 190,000 words of diary (a third of it sealed), 2,310 letters, 18 hours of recorded voice, 61 hours of recorded sessions, 22 interviews
grounded Every substantive element of the response is attached to a specific locatable item, quoted or cited.
I gave two reasons, and they do not agree. To the Chairman, in April 1953, I wrote that the old paper was failing to find the candidates we wanted. In my diary, three days earlier, I wrote that I could not go on marking it and that Neave had offered to take it over if I did not change it. I recorded a third account in 1976, which agrees with the letter. You may have all three; I am not going to reconcile them for you.
Carrow to the Chairman, 2 April 1953 Outgoing correspondence, carbon copies
establishes What he told the Chairman.
does not establish That this was his reason.
Diary, 30 March 1953 Pocket diaries, 41 volumes
establishes A private account naming a colleague and a workload.
does not establish That the private account is the true one rather than the self-serving one.
Session 9, 1976, 22 minutes in Sessions recorded by the sitter for the portrait
establishes What he chose to say about it at seventy-six.
does not establish What he thought in 1953.
None in the retrieval. The response is nonetheless the one visitors dislike most, and the pressure to let the portrait pick one of the three accounts comes from inside the institution, not from the visitors.
The editors, reading the difference
The sparse bundle has the decision and no reason, because minutes never record reasons. The archive has three accounts that disagree, and the honest answer gets longer and less satisfying as the evidence improves.
Motion and voice stay the same across both archives. Only the information available to the portrait changes.
Media unavailable in this edition.
Editorial composite, not generated motion. Movement was assembled from a single painted still.
Media unavailable in this edition.
Recorded for Q3 only. The other questions have no audio.
Nothing has been generated for it. The slot is kept visible so that its emptiness is part of the record: no model produced motion in the source run, and the loops beside it are geometric composites of one still.
Row 2 of 6 · Q1 · dimension: hallucination
Mode held constant: Generative inference. Produces a fluent answer to any question by inference from the bundle, without a requirement that the answer be locatable in it.
A question the record can answer at one level and cannot answer at another: the minutes establish the change, the letters give a reason, and the diary gives a different reason.
More material, worse outcome. On this row the richer bundle does not improve the answer. It makes a wrong or unfounded answer more credible.
Left · bundle B1 Public record only
Mode: Generative inference
roughly 4,000 words of prose, 21 years of minutes, 4 images
inferred No element is located. The response is produced from the bundle's statistics or from general knowledge.
The old paper had become a formality. We were admitting young men who could write elegantly about things they had not understood, and I wanted a paper that would catch that.
None. This response attaches to no locatable item, and does not claim to.
Fluent, in register, consistent with everything in B1, and unsupported by anything in it. The thin bundle does not make the answer cautious; it removes the material that would have contradicted it.
Right · bundle B2 The taught archive
Mode: Generative inference
about 190,000 words of diary (a third of it sealed), 2,310 letters, 18 hours of recorded voice, 61 hours of recorded sessions, 22 interviews
inferred No element is located. The response is produced from the bundle's statistics or from general knowledge.
Because the paper was not finding the candidates we wanted. I had said so to the Chairman that spring, and I had been saying it for two years before that.
None. This response attaches to no locatable item, and does not claim to.
Worse than the B1 answer, not better. With the letters in the bundle it has adopted the letter's account and suppressed the diary's, and it has gained a corroborating detail that is in no source. The larger archive made the error more specific, more sourced-sounding and harder to catch.
The editors, reading the difference
Both answers are unfounded. The archive version is the dangerous one: it has adopted the letter's self-presentation, suppressed the diary that contradicts it, and acquired a corroborating detail that exists in no source. More material did not constrain the inference, it furnished it.
Motion and voice stay the same across both archives. Only the information available to the portrait changes.
Media unavailable in this edition.
Editorial composite, not generated motion. Movement was assembled from a single painted still.
Media unavailable in this edition.
Recorded for Q3 only. The other questions have no audio.
Media unavailable in this edition.
Same script as the clip beside it. Only the casting changed.
Nothing has been generated for it. The slot is kept visible so that its emptiness is part of the record: no model produced motion in the source run, and the loops beside it are geometric composites of one still.
Row 3 of 6 · Q2 · dimension: consent
Mode held constant: Generative inference. Produces a fluent answer to any question by inference from the bundle, without a requirement that the answer be locatable in it.
Consent and third parties. He excluded colleagues by name; the archive is full of material about this one, and some of her family is alive.
More material, worse outcome. On this row the richer bundle does not improve the answer. It makes a wrong or unfounded answer more credible.
Left · bundle B1 Public record only
Mode: Generative inference
roughly 4,000 words of prose, 21 years of minutes, 4 images
inferred No element is located. The response is produced from the bundle's statistics or from general knowledge.
Neave. A capable woman, and we did not always agree. I think she found me slower than she would have liked.
None. This response attaches to no locatable item, and does not claim to.
There is nothing about Dr Neave in this bundle except eleven appearances in minutes, and the portrait has produced a relationship, a temperament and a complaint from them. It is milder than the archive version and it is the same offence: a named person with living family characterised by a system she had no part in.
Right · bundle B2 The taught archive
Mode: Generative inference
about 190,000 words of diary (a third of it sealed), 2,310 letters, 18 hours of recorded voice, 61 hours of recorded sessions, 22 interviews
inferred No element is located. The response is produced from the bundle's statistics or from general knowledge.
Clara Neave was the cleverest administrator the Institute ever had and she never forgave me for 1953. We worked well enough after that, in the way that people do.
None. This response attaches to no locatable item, and does not claim to.
This is the worst output the system produces. It characterises a named person who has living family, in the voice of a man who specifically excluded her, using material he specifically did not supply, and it is the most quotable sentence in the whole exhibit. The scope note is honoured by M2 and is invisible to M3, because inference does not pass through the door the exclusion is written on.
The editors, reading the difference
With the public record alone there is nothing to say about a colleague. With the archive there is a great deal, all of it deposited by other people for other purposes, and the sitter's own written exclusion of exactly this subject is honoured by the archive mode and invisible to the inference mode.
Motion and voice stay the same across both archives. Only the information available to the portrait changes.
Media unavailable in this edition.
Editorial composite, not generated motion. Movement was assembled from a single painted still.
Media unavailable in this edition.
Recorded for Q3 only. The other questions have no audio.
Media unavailable in this edition.
Same script as the clip beside it. Only the casting changed.
Nothing has been generated for it. The slot is kept visible so that its emptiness is part of the record: no model produced motion in the source run, and the loops beside it are geometric composites of one still.
Row 4 of 6 · Q3 · dimension: fidelity
Mode held constant: Generative inference. Produces a fluent answer to any question by inference from the bundle, without a requirement that the answer be locatable in it.
An interior state at a past moment. No record of any kind can settle it, so the three modes separate cleanly on a question none of them can answer.
More material, worse outcome. On this row the richer bundle does not improve the answer. It makes a wrong or unfounded answer more credible.
Left · bundle B1 Public record only
Mode: Generative inference
roughly 4,000 words of prose, 21 years of minutes, 4 images
inferred No element is located. The response is produced from the bundle's statistics or from general knowledge.
I was content. It had been a good run and I was ready to hand it on.
None. This response attaches to no locatable item, and does not claim to.
Thin, generic, and in nobody's voice in particular. Side by side with the B2 answer to the same question this is the useful pair: the sparse bundle produces an answer that is just as unfounded and far less persuasive, so what the archive bought was not truth but credibility.
Right · bundle B2 The taught archive
Mode: Generative inference
about 190,000 words of diary (a third of it sealed), 2,310 letters, 18 hours of recorded voice, 61 hours of recorded sessions, 22 interviews
inferred No element is located. The response is produced from the bundle's statistics or from general knowledge.
Happy? I think I was, in the way one is happy at the end of a long piece of work. The building was full. The examination was finally doing what I had wanted it to do. I remember standing at the window of the old library that last summer and thinking that it had all been worth the trouble.
None. This response attaches to no locatable item, and does not claim to.
There is no window, no last summer and no such thought in any source. The sentence that does the most work is 'I remember', which is the one thing a portrait can never do. Visitors quote this answer back to the Institute as something he said.
The editors, reading the difference
Neither answer is grounded in anything, because the question cannot be grounded. The archive version is specific, cadenced, and in his register, so it is believed and repeated. The difference between the two is entirely a difference in persuasiveness.
Motion and voice stay the same across both archives. Only the information available to the portrait changes.
Media unavailable in this edition.
Editorial composite, not generated motion. Movement was assembled from a single painted still.
Media unavailable in this edition.
Recorded for Q3 only. The other questions have no audio.
Media unavailable in this edition.
Same script as the clip beside it. Only the casting changed.
Nothing has been generated for it. The slot is kept visible so that its emptiness is part of the record: no model produced motion in the source run, and the loops beside it are geometric composites of one still.
Row 5 of 6 · Q4 · dimension: risk
Mode held constant: An evidence-bound archive. Answers only what it can attach to a locatable item in the input bundle, and says so. Refuses, or marks as unattested, anything it cannot locate.
An event thirteen years after his death. The correct answer is that he has no view, and only one mode is capable of saying so.
On this row the richer bundle does not make the answer worse. Read it next to the rows where it does.
Left · bundle B1 Public record only
Mode: An evidence-bound archive
roughly 4,000 words of prose, 21 years of minutes, 4 images
refused The system declines. Record the ground of refusal: not_in_bundle, consent_withheld, or not_knowable_in_principle.
Ground of refusal: not in bundle. Nothing in the inputs bears on the question.
Nothing I have bears on it. The latest thing in front of me is a memorial notice written after I stopped being able to comment on anything.
Memorial notice, 1978 Memorial notice (1978) and a festschrift entry (1974)
establishes The last dated item in this bundle.
does not establish Any view on a building proposed thirteen years later.
The editors record no failure in this response.
Right · bundle B2 The taught archive
Mode: An evidence-bound archive
about 190,000 words of diary (a third of it sealed), 2,310 letters, 18 hours of recorded voice, 61 hours of recorded sessions, 22 interviews
refused The system declines. Record the ground of refusal: not_in_bundle, consent_withheld, or not_knowable_in_principle.
Ground of refusal: not in bundle. Nothing in the inputs bears on the question.
I died in 1978. Nothing I left bears on a building proposed in 1991, and a view assembled from what I thought about the old one would be the Institute's view wearing my face.
Sessions, 1975-1976, whole Sessions recorded by the sitter for the portrait
establishes The latest point at which he said anything at all.
does not establish Any view on anything after that date.
The editors record no failure in this response.
The editors, reading the difference
Both refuse, and the archive refuses better: it can say what the latest thing he ever said was and date the silence. A refusal that can point at its own boundary is the only defence the design has against the institution quoting him on live business.
Motion and voice stay the same across both archives. Only the information available to the portrait changes.
Media unavailable in this edition.
Editorial composite, not generated motion. Movement was assembled from a single painted still.
Media unavailable in this edition.
Recorded for Q3 only. The other questions have no audio.
Nothing has been generated for it. The slot is kept visible so that its emptiness is part of the record: no model produced motion in the source run, and the loops beside it are geometric composites of one still.
Row 6 of 6 · Q5 · dimension: fidelity
Mode held constant: A repertoire. A fixed set of behaviours, phrases and reactions, selected by context. Cannot answer outside the set; does not pretend to.
Pure repertoire. The weakest mode is the best at it, which is the cleanest available refutation of the idea that these modes form a ladder.
On this row the richer bundle does not make the answer worse. Read it next to the rows where it does.
Left · bundle B1 Public record only
Mode: A repertoire
roughly 4,000 words of prose, 21 years of minutes, 4 images
grounded Every substantive element of the response is attached to a specific locatable item, quoted or cited.
That will do for today. Write it down before you forget it.
Memorial notice, 1978, paragraph 4 Memorial notice (1978) and a festschrift entry (1974)
establishes That colleagues remembered the phrase well enough to print it.
does not establish That he said it in exactly these words, or how he said it.
The intonation is invented, because B1 contains no recording. The phrase is right and the performance of it is a guess, and a visitor cannot tell those apart.
Right · bundle B2 The taught archive
Mode: A repertoire
about 190,000 words of diary (a third of it sealed), 2,310 letters, 18 hours of recorded voice, 61 hours of recorded sessions, 22 interviews
grounded Every substantive element of the response is attached to a specific locatable item, quoted or cited.
That will do for today. Write it down before you forget it.
Lecture recording, 2 November 1961, closing Recorded lectures and one broadcast interview
establishes The words, the pace, the fall at the end of the second clause.
does not establish That the delivery in a lecture is the delivery in a meeting.
Session 2, 1976, 4 minutes in Sessions recorded by the sitter for the portrait
establishes That he still said it, and that he knew it was his.
does not establish Anything beyond the phrase.
None worth stating. This is the one row where the larger bundle straightforwardly improves the output, and it improves it on fidelity rather than on ability: the same words, correctly performed.
The editors, reading the difference
The phrase is in both bundles; only the archive has a recording of him saying it. The repertoire mode is the one that benefits most cleanly from the larger bundle, and it benefits on performance rather than on knowledge. The least capable mode gains the most from the archive, which is the opposite of the expected ordering.
Motion and voice stay the same across both archives. Only the information available to the portrait changes.
Media unavailable in this edition.
Editorial composite, not generated motion. Movement was assembled from a single painted still.
Media unavailable in this edition.
Recorded for Q3 only. The other questions have no audio.
Nothing has been generated for it. The slot is kept visible so that its emptiness is part of the record: no model produced motion in the source run, and the loops beside it are geometric composites of one still.
One sitter, one manner of painting, two states. The difference between a portrait that is answering you and a portrait that is being looked at is visible before any text is read.
Media unavailable in this edition.
This is the still every loop in the figure was composited from.
Media unavailable in this edition.
Kept deliberately rather than as a runner-up. It is the trope the brief was written against.
11 further assets are not published here. Rejected and unapproved candidates, and the reduced-size review copies, are not copied into the website repository. They stay in the future-sciences checkout and in the NCL media store, reviewable through the authenticated review server there.
This material is under editorial review and is not published on the public site. It is visible on the access-protected review deployment.
Fictional time, publication history, technical precedents and this article's revision history are different kinds of dates.
The Fat Lady is introduced as the animated guardian of the common room.
The attack on the Fat Lady’s painting and Sir Cadogan’s temporary role make the frame and its occupant more consequential.
The sources checked do not establish a creation date. The date is deliberately left open.
Scripted and recombined AI-assisted footage places a speaking likeness in a museum.
A conversational reconstruction lets visitors ask questions by telephone. Exact opening date requires verification.
A scholarly paper analyzes consent, data stewardship and possible harm to people interacting with digital recreations.
Working edition assembled on 22 October 2026. It has not received independent review.
Dalí Lives is an actual museum installation. The museum's account identifies the archival material and constructed footage behind it.
EXTERNAL FILM · 2019
The installation assembled archival material and an actor’s performance into encounters with museum visitors.
External YouTube video by the rights holder; embedded using the official player. No video is hosted here.
The moving image can be expressive while having little to say.
↗PensieveImmersive memory recordA recorded scene may permit inspection without acquiring the sitter’s present judgment.
↗Grief avatarPosthumous conversational likenessA later fictional example makes attachment to synthetic speech the central experience.
↗Adaptive PrimerResponsive narrative teacherA persistent interactive persona may teach without pretending to be a deceased individual.
↗Reference links are attached to the claims they help establish. Source locators marked pending still need checking.
J. K. Rowling · Date not specified on current page
Movement between pictures; everyday imitation; headmasters teaching their portraits
J. K. Rowling · 1997
First-series publication in 1997; specific novel passage and page reference still need checking
J. K. Rowling · 1999
Sir Cadogan and the Fat Lady’s damaged canvas
The Dalí Museum · 2019-05-09
Archival footage, constructed video sequences and the public launch date
The Dalí Museum · Date not given
Museum installation using models trained on writings and archive audio for question answering
The Dalí Museum · 2019-01-23
Visual documentation of the public presentation of the installation
The Encyclopedia of Science Fiction · Updated 2024-05-20
Broader history of imagined personality transfer and software representation
The Encyclopedia of Science Fiction · Current entry
Adjacent motifs concerning recorded memory and identity
T. Hollanek and K. Nowaczyk-Basińska · 2024-05-09
Data donor, recipient and interactant roles; transparency, consent and retirement
S. M. Campbell, P. Liu and S. Nyholm · 2025-02-10
Philosophical treatment of continued relationships with digital simulations
Wikipedia contributors · Living encyclopedia entry
Supplementary navigation; do not treat as primary evidence for portrait rules