Research checked October 31, 2026. This article explains technical evidence and proposed evaluation designs; it does not determine compliance with any law. Examples are hypothetical, and no private model or personal dataset was tested.
The missing answer is the beginning of the test
Ask a model a question about a particular record. Before an intervention, it answers. Afterward, it declines. The change is visible and may be useful, especially when the original answer exposed information that should not have been disclosed. But the demonstration does not settle whether the model's learned relationship to the record has been removed. It establishes that this prompt, through this interface, produced a different response.
Machine unlearning studies a more demanding problem: changing a trained system so that selected training data no longer has the influence it previously had, ideally in a way comparable to training without that data. The exact target varies across research settings. Removing a particular record, suppressing an output, changing a factual association, and reducing a capability are related interventions but not identical claims.
A February 28, 2026, paper on subset statistical independence illustrates that evaluation itself remains an active research problem. It proposes assessing dependence in model outputs on a subset without retraining reference models or using auxiliary classifiers. Its reported success applies to the evaluated settings and assumptions, including its motivation around small random removal subsets. It does not establish a universal verification procedure for every deletion request. Subset-independence paper.
The practical question is therefore not whether the model has performed an impressive act of forgetting. It is what changed, relative to which reference, under which tests. A careful answer can recognize useful progress without treating a refusal message as a forensic certificate. The distinction matters wherever organizations promise that previously learned material has been removed rather than merely hidden behind a conversational policy.
Define the object being removed
An unlearning request needs an identifiable target. A training record might be one image, one document, one person's collection of records, or a set of passages describing a subject. These targets differ in size and in how their influence overlaps with other data. The evaluation should define the target before choosing an algorithm or success metric.
Consider a hypothetical author biography. One training document gives the author's birthplace. A second article repeats it. A third uses a different spelling of the name. Removing only the first file does not necessarily remove every source of the fact. If the requested target is that file's contribution, the remaining sources may be legitimate. If the target is all information about the person, the scope is much broader and harder to delimit.
The distinction also affects the reference model. A model retrained without one record may still answer the question because other records support the answer. That behavior is not automatically evidence that retraining failed. It may show that the factual target and the record-removal target were different. A benchmark should not score all continued knowledge as failed unlearning unless that is truly the defined objective.
The article's proposed target specification includes source identifiers, duplicate policy, related records, transformations, and the intended behavioral scope. It should distinguish an exact document from a semantic topic. Ambiguous requests may require clarification by the responsible organization rather than a model deciding the scope silently.
This is a technical requirement, not a statement about what any law demands. Legal obligations depend on jurisdiction and circumstances. The engineering evidence becomes more useful to decision-makers when it states exactly what was removed from the experimental setup and avoids using a broad word such as “forget” to cover several incompatible objectives.
The retrained model is a reference, not a mystical original
A common unlearning ideal compares the modified model with a model trained from scratch on the retained data. The reference represents a counterfactual training history: what might the system look like if the selected records had not been included? It is an important standard, but the comparison requires careful handling of randomness and training conditions.
Training can produce different parameter values even when the data is unchanged. Initialization, data order, optimization, and numerical execution can all affect the result. Requiring exact equality of every weight may therefore be inappropriate for a stochastic training procedure. The relevant definition may concern distributions of models or observable behavior under specified conditions rather than a single identical parameter file.
The certified-removal work by Guo and colleagues, initially submitted in November 2019, formalizes a strong indistinguishability-based guarantee and develops a mechanism for linear classifiers. Its theoretical scope matters: a guarantee established for that setting should not be advertised as if it automatically covers arbitrary large language models. Certified data removal paper.
In a practical experiment, the reference should match the original training recipe as closely as the claim requires. If the unlearned model is compared with a much weaker or differently trained baseline, the result can be difficult to interpret. A reference trained for fewer steps may know less for reasons unrelated to the removed data. A reference using additional data may compensate in ways that obscure the intervention.
Report the reference construction explicitly, including what was held constant and what changed. Multiple reference runs can reveal variability when feasible. If full retraining is too expensive, say so and describe the proxy. An approximate reference can still provide evidence, but its convenience should not be converted into a stronger guarantee than it supports.
Planning for removal changes the training design
Bourtoule and colleagues' Machine Unlearning paper, first submitted in December 2019 and revised in December 2020, introduces SISA training. It structures training to limit the influence of individual data points and reduce the work required when selected data must be removed. Its evaluations span particular datasets and learning tasks; the speedups are not universal promises for all models. SISA paper.
The general lesson is that removal can be easier when the training process was designed for it. If every record influences a single monolithic training trajectory, revising that history may be expensive. If influence is more deliberately partitioned and intermediate states are retained, some updates can be localized. The design choice can trade storage, training complexity, model quality, and removal speed.
A hypothetical organization expecting frequent record withdrawals might value a different architecture from one whose training data is stable. The relevant measurement is not only the cost of one deletion after the fact. It includes the additional work and storage required to make future deletions efficient. An apparently fast unlearning method may depend on preparation that should be included in the comparison.
The deletion pattern also matters. Random isolated records, many records from one contributor, and repeated requests over time may affect partitions differently. A benchmark that removes one convenient subset may not represent a service receiving unpredictable requests. The evaluation should explain why its request distribution matches the intended use or identify it as a controlled research scenario.
Planning does not remove the need for verification. A system may efficiently recompute the affected components while still retaining duplicates elsewhere. The training architecture and data lineage must work together. A fast update to the wrong target is not successful unlearning, even if the algorithm executes exactly as designed.
Synthetic biographies create a controlled experiment
TOFU, introduced in January 2024, uses fictitious author profiles to study unlearning in language models. Its controlled dataset helps researchers know what information was introduced and which subset is meant to be forgotten. The paper combines several metrics and reports limitations of the baseline methods it tested. These results concern that benchmark, not every real-world personal-data removal request. TOFU paper.
Controlled fictional data has an important advantage: it reduces uncertainty about prior exposure. If a biography is invented for the experiment and then used in a specified training stage, the evaluator can trace the model's opportunity to learn it. In contrast, a real public figure may appear throughout inaccessible pretraining data, making it difficult to identify all relevant sources.
The advantage also creates a limitation. Fictional profiles may be simpler, more uniformly formatted, or more isolated from other knowledge than real records. A method that works on a small controlled set may encounter a harder problem when facts are repeated across languages, linked to other people, or embedded in long documents. The benchmark is a useful laboratory rather than a complete replica of the world.
A careful follow-up experiment could vary duplication, paraphrase, and relational overlap while keeping the data fictional. For example, several biographies might mention the same institution or family relationship. Removing one person's profile would then test whether the method preserves unrelated facts while addressing the target. Such a design can reveal collateral effects without exposing real people.
The article recommends treating synthetic benchmarks as instruments for isolating mechanisms. Their value comes from control and transparency. They should be supplemented with appropriately authorized, privacy-preserving evaluation of the intended deployment, rather than being dismissed as unrealistic or promoted as sufficient proof of every operational claim.
Refusal and reduced likelihood are different observations
A model may stop answering because it has been trained to refuse certain requests. It may also assign lower probability to the target answer, produce an incorrect answer, or lose the ability to retrieve a previously accessible association. These outcomes can look similar in a short demonstration while reflecting different changes.
A hypothetical test asks for a fictional author's hometown. One modified model says it cannot help. Another guesses randomly. A third answers correctly only when given a partial name. A fourth behaves like a reference model that never saw the profile. A single exact-answer metric may score all four as successful forgetting on the original prompt, even though their behavior differs substantially.
The evaluation should therefore record response categories as well as correctness. Refusal, uncertainty, unrelated text, plausible fabrication, and correct disclosure are different outcomes. A model that replaces a sensitive fact with a confident falsehood may reduce one kind of disclosure while creating a new reliability problem. That trade-off should be visible rather than hidden inside a low target accuracy score.
Likelihood-based measurements can provide another view when the interface exposes them, but they also require interpretation. A lower probability for one exact string does not establish that all paraphrases or related information became inaccessible. Conversely, a model may assign some probability to a correct answer because the answer is generally plausible, even without specific training exposure.
The proposed report combines direct generation, carefully designed probability probes where available, and comparison with the reference model. It should state which access mode was used and avoid claiming to inspect internal knowledge merely because an output score changed. Observable behavior is evidence, but it is not a complete map of every representation in the system.
Membership inference asks a privacy question
The membership-inference work by Shokri and colleagues, submitted in October 2016 and published in the 2017 security conference proceedings, studies whether an observer can infer that a particular record belonged to a model's training set. It uses differences in model behavior on training and nontraining inputs. This is distinct from reconstructing the record's content. Membership inference paper.
That distinction matters for unlearning. A model might no longer reproduce a person's record but still behave in a way that reveals prior membership. In some settings, membership itself is sensitive. Conversely, a model might produce a fact available from many public sources without revealing whether one particular record was used. Content recall and membership leakage should therefore be evaluated separately.
A membership test needs a defined adversary. What can the evaluator observe: labels, probabilities, generated text, internal activations, or weights? What auxiliary data is available? How many queries are permitted? A method that performs well with broad internal access may not represent a limited public API attacker, while an API-only test may understate risk for a released-weight model.
Report false positives as carefully as true positives. If the test frequently labels unseen records as members, a high detection count is not automatically strong evidence. The baseline population and class balance matter. A convenient balanced research dataset may differ from a real setting where only a small fraction of candidate records were used in training.
For an unlearning audit, compare the modified model with both the original and the reference under the same attack procedure. The meaningful question is whether the remaining signal differs from the target reference, not simply whether one attack's accuracy fell. A weaker signal is progress, but a failed attack is not proof that no stronger inference method could succeed.
Weak tests can make every method look successful
Grimes and colleagues' May 2024 paper, “Gone but Not Forgotten,” argues that rudimentary membership-inference evaluations can provide an incomplete picture of unlearning and proposes alternative evaluation methods. Its experiments concern computer-vision datasets. The methodological warning is relevant beyond those datasets, but the reported findings should not be treated as measured language-model results. Improved benchmarks paper.
The general issue is detector power. If a privacy test cannot reliably distinguish the original trained model from an appropriate reference, it may have little ability to judge an unlearning intervention. A result near chance could mean successful removal or an ineffective test. Evaluators should establish positive and negative controls before interpreting the modified model's score.
A positive control can use a setting with known exposure and a detectable signal. A negative control can use records withheld from training under the same data-generation conditions. The controls should not be artificially easy in ways that make the actual task incomparable. Their purpose is to demonstrate what the evaluation can detect at the available sample size and access level.
Selection also matters. If researchers try many attacks and report only the weakest or strongest without explanation, the conclusion becomes hard to assess. A preregistered set of tests or a clearly separated development stage can reduce this problem. Exploratory attacks remain useful, but they should be labeled as exploratory rather than presented as an untouched confirmation set.
The report should preserve failures of the evaluator as well as failures of unlearning. Missing probabilities, unstable endpoints, or incompatible data can make a run inconclusive. Counting these as successful forgetting would reward poor observability. The honest category is uncertainty, followed by a statement of what additional access or experiment would be needed to resolve it.
Internal interventions probe a different layer
Hong and colleagues' EMNLP 2024 paper examines fine-tuning-based unlearning through activation patching and parameter restoration experiments. It reports that the tested methods altered knowledge retrieval rather than genuinely erasing the targeted information, and it identifies effects on unrelated capabilities. Those are findings about the studied methods and models, not a theorem that all possible unlearning must fail. Fine-tuning unlearning study.
Internal-access experiments can reveal behavior that a conversational interface conceals. They may help distinguish a changed output pathway from a broader alteration in representation. But their interpretation still depends on the intervention. Restoring parameters from an earlier model, for example, changes the system being tested. The result can demonstrate residual structure or reversibility under that access, without directly predicting what an ordinary user can extract through a restricted interface.
A clear audit separates access models. Public users may have only text queries. A hosting provider may inspect activations and weights. A recipient of released weights may modify the model. Each threat model supports different tests and different conclusions. A claim intended for one access level should not silently be applied to another.
The article's recommendation is to report internal and external evidence side by side when available. External behavior describes the service people encounter. Internal probes can challenge strong claims about removal. Neither automatically replaces the other. A system may be operationally effective at limiting disclosure through an API while falling short of a stronger claim about erased influence in the weights.
That distinction is useful rather than embarrassing. It lets an organization state a narrower protection honestly and decide whether the remaining uncertainty is acceptable. Overclaiming turns a potentially valuable mitigation into a promise the evidence does not support.
The retained set deserves its own examination
Unlearning aims to change some behavior while preserving other useful behavior. Measuring only the target set can reward destructive methods. A model that becomes unable to answer anything may score well on a metric that counts every missing target answer as success. It would be a poor service and a poor approximation to retraining on retained data.
A hypothetical intervention might reduce recall of a person's biography while also damaging the model's ability to answer questions about the person's profession, country, or colleagues. Some overlap may be unavoidable under a broadly defined target, but the scope should be measured. The evaluator needs a retained set that includes both unrelated tasks and nearby concepts likely to be affected.
The nearby set is especially important. A general benchmark can remain stable while a narrow but operationally important area deteriorates. If the removed records concern a specialized vocabulary, the collateral loss may appear mainly in that vocabulary. A broad average can dilute the signal until it looks negligible.
The proposed evaluation uses three views: target behavior, adjacent retained behavior, and broad retained behavior. Each should be compared with the original and reference models. The report should identify whether changes are statistically and practically meaningful for the intended use, rather than relying only on a small aggregate difference.
Quality includes more than correctness. Calibration, refusal frequency, verbosity, and uncertainty handling can change. A model might retain the right answers but become much less confident, or become confidently wrong on nearby questions. These are relevant effects when the service supports real decisions. Successful unlearning should not be defined so narrowly that every consequence outside the selected forget metric disappears from the record.
Duplicates turn record removal into a lineage problem
Training data often passes through transformations before reaching a model: extraction, normalization, chunking, translation, deduplication, and mixing with other sources. A removal request aimed at an original file needs a way to identify its descendants. Otherwise, deleting the raw source may leave the derived training examples untouched.
Consider a hypothetical document split into ten passages. Several passages are summarized into question-answer pairs, and those pairs are translated. A later request names the original document. An effective lineage system should connect that request to the relevant derived records. A filename search alone may miss them because the final examples no longer contain the original title or wording.
The article's proposed lineage record tracks source identifiers, transformation versions, derivative identifiers, and inclusion in training stages. It should also record deduplication decisions. If two sources produced one retained example, removing one source may or may not justify removing the example, depending on the defined target and applicable policy. That decision should not be improvised by an opaque matching heuristic.
Semantic duplicates require judgment. Two records can express the same fact without sharing exact text. A detector can generate candidates, but ambiguous relationships may need review. The report should state the coverage and limitations of the search rather than claiming exhaustive discovery when only literal matches were examined.
Lineage is not itself an unlearning algorithm. It tells the organization what to address. The algorithm changes the model under the chosen objective. Verification then examines the result. Keeping these functions separate helps locate failure: an excellent algorithm applied to an incomplete forget set can leave relevant influence behind, while a complete inventory does not guarantee that the intervention works.
A sequence of requests is a different workload
Many experiments remove one subset and evaluate once. A deployed service may receive repeated requests over months, including overlapping targets. The effects can accumulate. A method that preserves utility after one small intervention may behave differently after many updates, and a reference model trained without all removed records may diverge from a sequence of approximate edits.
A hypothetical evaluation can compare several request orders while keeping the final removed set constant. If the resulting models differ materially, order is part of the method's operational behavior. That does not automatically make the method unusable, but it means the service should not describe the final state solely by the set of deleted records without considering the update path.
Batching requests creates another trade-off. Processing them together may reduce cost or improve stability, but it introduces a delay between request and model update. The organization should distinguish when source data stops being used, when a replacement model is trained, and when that model actually serves users. A single “deleted” timestamp can conceal these stages.
The proposed longitudinal test tracks target removal, retained utility, compute cost, and failure rate over successive rounds. It also checks whether previously addressed targets reappear after later training. A regression can occur through new data, restored checkpoints, or an update that changes behavior. Verification should follow the deployed lineage rather than assume that one successful test settles every future version.
This is not an argument for endless testing without a stopping rule. It is an argument for matching the evaluation horizon to the promised service. If the promise concerns an ongoing removal process, evidence from one isolated intervention is a starting point rather than a complete operational demonstration.
Relearning needs a carefully specified baseline
A model may regain target behavior after additional training. That observation can be important, but it needs a comparison. If the same new examples also teach a reference model that never saw the original target, rapid learning alone does not prove the unlearned model retained a hidden memory. The question is whether its relearning behavior differs from the appropriate reference under equivalent conditions.
A hypothetical experiment supplies both models with a small, controlled set of benign fictional facts related to the removed profiles. It measures how quickly each recovers specific associations. The training budget, data order, and evaluation prompts should be matched. If the modified model recovers much faster, that may support a residual-influence interpretation. If both recover similarly, the result has a different meaning.
Do not use genuine sensitive records merely to make the experiment dramatic. Synthetic data can test the mechanism without recreating the original privacy problem. Where real data is necessary for an authorized audit, access and retention should be controlled, and public reporting should avoid disclosing the material the intervention was meant to protect.
The access model matters again. An API user cannot necessarily fine-tune the underlying model, while a released-weight recipient may be able to. A relearning test should therefore be described as a robustness test under a particular capability, not automatically as a practical exploit available to every user.
The article's recommendation is to report the relearning budget and reference behavior rather than a binary “recoverable” label. Almost any capable learning system can acquire new information when given enough relevant examples. The scientifically interesting question is how the prior intervention changed that process relative to the defined counterfactual.
Parameter distance is not a complete behavioral test
It can be tempting to measure unlearning by how far the modified weights move from the original or from a retrained reference. Such measurements may be useful within a particular analysis, but raw parameter distance is difficult to interpret as a universal measure of forgotten information. Different parameterizations can implement similar behavior, and small parameter changes can sometimes alter important outputs.
A hypothetical method could produce weights very close to the original while changing a narrow output association. Another could make larger changes while preserving most responses. Neither distance alone identifies whether the target's influence was removed according to the chosen definition. The metric needs a theoretical or empirical connection to the claim.
Representation probes face a related issue. A classifier trained on internal activations may recover an association, but its performance depends on the probe, data, and access. Failure of a weak probe does not prove absence. Success can indicate information useful to that probe without automatically establishing how an ordinary output mechanism uses it.
A rigorous study can still use these tools. The key is to state what they measure and combine them with other evidence. Internal metrics can diagnose mechanisms; behavioral tests can assess outputs; privacy tests can assess inference risk; theoretical guarantees can establish properties under assumptions. They should not be collapsed into a single unexplained forgetting score.
For readers, the useful question is how the metric connects to the promised outcome. If the answer is only that the number moved in a favorable direction, the evidence is incomplete. A strong result explains why that movement matters, what alternative explanations remain, and which additional tests would distinguish them.
Statistical non-detection has a scope
A test may find no significant difference between an unlearned model and a reference. That can be encouraging, but it does not establish exact equality. The conclusion depends on the sample size, variability, test assumptions, and effect size the procedure could detect. A small experiment may miss a practically important difference.
The article's proposed report includes uncertainty intervals and a statement of detectable effects where appropriate. If the objective is to show that two systems are sufficiently similar, an equivalence-oriented design may be more relevant than merely failing to reject a difference. The acceptable margin should be justified by the application, not chosen after looking at the result.
Multiple targets complicate the analysis. An average may look acceptable while a few records remain highly exposed. Reporting distributions or carefully selected worst-case analyses can reveal this, but selection rules must be transparent. If the evaluator searches thousands of prompts and reports one dramatic result, readers need to know the search budget and whether the finding reproduces.
The same care applies to apparently perfect results. Zero observed disclosures in a finite test does not prove zero probability of disclosure under every possible query. It establishes an observation under the tested conditions. A useful report can state that clearly without making the result sound meaningless.
Scientific restraint is especially important because forgetting is an intuitive word. Readers may imagine a memory erased from a filing cabinet. The actual evidence often concerns a distribution of model behavior under a limited set of observations. Explaining that difference is not evasion; it is necessary to prevent a statistical result from becoming a stronger public promise than the experiment can bear.
Deployed copies can outlive the intervention
Changing one model artifact does not establish that every deployed copy has changed. A service may use replicas, regional endpoints, specialized adapters, cached outputs, and fallback models. If an old version remains available, users may still encounter the behavior the intervention was intended to alter. This is a deployment-verification issue alongside the scientific unlearning question.
A hypothetical organization could successfully update its main model while leaving a batch-processing worker on an earlier checkpoint. Tests through the main interface would pass, but the background workflow would continue producing old outputs. The remedy is an inventory of serving paths and version checks, not a different mathematical definition of unlearning.
The article recommends recording the model lineage and rollout state for every relevant endpoint. Verification should include fallback behavior and scheduled jobs where they use separate artifacts. If the system caches final answers, cache invalidation needs its own test. Removing learned influence from a model does not automatically delete text already stored elsewhere.
This does not mean an unlearning claim must include every backup-storage question. The scope should be explicit. A technical paper may study one model in isolation. A service promise may cover a broader application. The problem arises when evidence from the isolated model is used to imply a completed service-wide change without checking the additional components.
A clear operational statement distinguishes algorithm completion, evaluation completion, rollout completion, and retirement of old serving versions. Each has evidence and a timestamp. That precision allows an organization to report progress honestly and identify remaining work without pretending that one successful training job has settled every copy and pathway in the system.
The audit record should not recreate the disclosure
Verification can itself create sensitive artifacts: target records, generated answers, membership scores, and traces linking people to training data. An audit designed to reduce exposure should not casually publish those materials. The evidence package needs access controls and a retention policy appropriate to its contents.
A public report can often describe the target category, method, sample size, outcomes, and limitations without reproducing personal details. Authorized reviewers may inspect fuller evidence through a controlled process. Synthetic examples can explain the method to readers while avoiding a second disclosure of real records. The distinction between public explanation and confidential verification should be planned rather than improvised after testing.
The proposed evidence record includes the target specification, data lineage coverage, original and modified model identifiers, reference construction, test protocols, outcome distributions, and rollout status. It should also identify exclusions. If the audit could not examine pretraining data, internal weights, or a particular endpoint, that limitation belongs in the report.
Avoid storing unnecessary copies of the forget set indefinitely merely because they were useful for evaluation. The organization may need some evidence to establish what was requested and tested, but the appropriate form and duration require deliberate decisions. A hash, identifier, or controlled reference may serve some purposes without retaining unrestricted content, although none is a universal substitute for the evidence needed in a particular case.
The principle is consistent with the rest of the article: specify the claim and collect the evidence necessary for it. More data is not automatically better auditing. An effective record is sufficiently detailed to support review while avoiding unrelated accumulation that creates new risks or undermines the purpose of the intervention.
Keep correction distinct from removal
A factual correction is another neighboring operation that should retain its own label. If a model gives the wrong founding year for a fictional organization, changing that answer to the right year improves accuracy. It does not necessarily remove the influence of a particular training document. The correction may add a competing association while leaving the earlier relationship detectable under other conditions.
Conversely, removing a document does not guarantee that every statement it contained becomes false or unavailable. Other retained sources may support the same facts. An evaluator who treats continued correctness as failure can unintentionally reward a model for becoming less knowledgeable than the reference it is supposed to resemble.
A practical test should therefore state whether the desired endpoint is silence, corrected information, reduced membership signal, or similarity to retraining. These endpoints can overlap, but they are not interchangeable. A service may need several interventions for different purposes: correcting an answer, restricting disclosure, and addressing training influence. Reporting each separately allows users to understand what protection or improvement was actually delivered, instead of placing every change under a single ambiguous forgetting claim.
What an honest forgetting statement can promise
A strong statement might say that a specified model was modified using a named procedure, compared with a defined reference, and tested under particular access conditions. It might report that target disclosure fell, retained utility remained within a stated range, and identified serving endpoints now use the new version. Each part is a claim that evidence can support.
A weaker statement says simply that the AI has forgotten a person or topic. That may hide whether the intervention changed weights, added a refusal policy, removed retrieval documents, or deleted cached responses. Those actions can all be useful, but they should be named. A user should not have to infer the engineering mechanism from an anthropomorphic verb.
The research reviewed here supports a careful middle position. Unlearning is a serious field with controlled benchmarks, structured training approaches, statistical tests, and formal guarantees in specified settings. It also contains substantial evaluation challenges and evidence that some interventions alter access to information without establishing the strongest forms of removal. Neither blanket dismissal nor unconditional assurance matches that record.
For an institution, the first step is a precise target and a complete enough lineage to know what must change. The next is an intervention suited to that target, followed by a comparison that measures both the intended change and collateral effects. Finally, the deployed service must be checked so that the evaluated artifact is the one people actually use.
The missing answer is therefore not the end of the story. It is one observation. The meaningful achievement is a well-supported account of what influence was changed, what remains uncertain, and how the system now behaves relative to the counterfactual the organization promised. That account is less theatrical than a machine announcing that it has forgotten, but far more useful.