The AI claims, examined in detail.

Marici has made specific, falsifiable, and unusually expansive claims about the AI it operates: a “world's most sophisticated” system, over 50 AI tools, predictive mapping “at 80%+ accuracy,” and named AI products that purport to do the work of prosecutors, behavioural scientists, and intelligence analysts. This page treats each of those claims in turn. For each, it asks a single question: what does the public record ordinarily contain for an organisation that has built what Marici says it has built — and what can be located for Marici?

This is not an investigation into whether Marici's tools work as described. That would require access to the tools themselves and is not possible from public materials. It is a comparison of artefacts: the documentation, infrastructure disclosures, peer review, named partnerships, engineering staffing, evaluation methodology, and audit trails that real AI organisations leave behind. Their absence is not proof of anything; it is, however, the question this site is asking Marici to answer.

A short glossary, since some of the terms below are technical: LLM — large language model; RAG — retrieval-augmented generation; IRB — institutional review board, governs research on human subjects; MOU — memorandum of understanding, the document that ordinarily governs data-sharing between a nonprofit and a government agency; NCMEC — the US National Center for Missing & Exploited Children; CSAM — child sexual abuse material; Daubert — the US standard for admissibility of expert evidence.

The phrase “world's most sophisticated” is a superlative. Superlatives in computing are normally measured, not asserted. When organisations like OpenAI, Anthropic, Google DeepMind, or Meta AI describe their systems as state of the art, they do so by reference to specific public benchmarks — MMLU, GPQA, HumanEval, SWE-bench, MATH, BIG-bench, and dozens of others — with results posted to leaderboards anyone can replicate. Their model architectures are described in technical reports; their training compute is disclosed in order of magnitude; their evaluation methodology is published; and third-party laboratories independently reproduce the headline numbers.

For a smaller laboratory making a comparable superlative claim, the public record typically contains at minimum: a named principal investigator with relevant publications; one or more peer-reviewed papers describing the system; a technical whitepaper or model card; a public benchmark result or independent evaluation; and a named institutional affiliation (university, established lab, or major industry research group) under which the claim is being made.

No technical whitepaper, no published benchmark, no peer-reviewed publication, no named model provider, no public technical report, no third-party laboratory citation, and no independent evaluation could be located as of the date of this update. The phrase “world's most sophisticated AI” does not appear to have been used about Marici by any party other than Marici itself.

Against what benchmark or comparator is the system measured? Where are the results published? Which laboratory has independently evaluated it? Has the comparison been made by any third party in print?

A product surface of 50 distinct AI tools is unusual at any scale. The closest comparable in the anti-trafficking sector is Thorn, which publishes engineering case studies, names its primary platform (Spotlight), names its model providers, names its law-enforcement partners, and lists its engineering staff publicly. Thorn's published engineering team numbers in the dozens; its annual technology spend appears in its IRS Form 990 in line items that are easily orders of magnitude larger than Marici's reported technology and software expenses.

For a US 501(c)(3) of Marici's audited size (FY2024 total expenses $5.48M, salaries and wages $206,552), a 50-tool product surface would ordinarily imply some combination of: a meaningful engineering headcount (a single competent ML engineer in the United States rarely costs less than $200,000 fully loaded); a vendor invoice trail (model APIs from OpenAI / Anthropic / Google / Cohere, cloud compute from AWS / GCP / Azure, vector or data infrastructure from Pinecone / Snowflake / Databricks); a public engineering presence (job postings, GitHub, technical blog, engineering staff identifiable on LinkedIn); and Form 990 expense lines that reflect the technology footprint.

Three of the 50 tools are named on the public homepage. Each is described in a single sentence. The remaining forty-seven are not inventoried. No public-facing engineering job posting could be located. No GitHub presence or technical blog could be located. No named technical staff are identified on the Marici leadership page or in publicly visible LinkedIn results filtered to the organisation. The Form 990 financial table on the claims page shows the audited revenue, expenses, and salary lines for 2017–2024; readers can compare those figures to what a 50-tool engineering operation would ordinarily cost.

None of this is evidence that the tools do not exist. It is evidence that the public artefacts ordinarily produced by an organisation operating a 50-tool AI surface could not be located.

Is the full inventory of the 50 tools available? Which of them are independently developed, and which are wrappers around third-party APIs? Which model providers are used? Where are the engineering staff named? What share of the audited annual expenses is allocated to technology development, and how is that line reconciled with a 50-tool product surface?

Accuracy is a measurement, not a property. To say a system is 80% accurate is to say four things: (1) the prediction target is defined (locations? individuals? networks? events? recurrence? what specifically is being predicted?); (2) a ground-truth dataset exists against which the prediction is scored; (3) a held-out evaluation set was used (not the data the model was trained on); and (4) the false-positive rate is known and disclosed.

In trafficking work, the false-positive rate is uniquely consequential. A predictive system that misidentifies a non-trafficking location, individual, or network at scale creates real harms: surveillance of innocent people, misdirection of law-enforcement resources, exposure of survivors to mistaken intervention. Published academic work in this domain (cf. peer-reviewed studies in IEEE conferences and the Journal of Human Trafficking) is unusually careful about false-positive disclosure for exactly this reason.

The training data needed to make a predictive-mapping claim at this scale is not held in open repositories. Trafficking-pattern data of sufficient resolution is concentrated with federal agencies (Department of Justice ICAC task forces, FBI Innocence Lost initiative, Homeland Security Investigations) and the National Center for Missing & Exploited Children. Access is restricted and is governed by named data-sharing agreements with named legal terms. A nonprofit that has trained a predictive model on this data substrate would, in the ordinary course, disclose the named federal partner that authorised the access.

No definition of what “mapping” is predicting could be located. No methodology document. No held-out evaluation set or confusion matrix. No false-positive rate disclosure. No named federal data-sharing partner. No IRB or ethics-board approval reference. No academic collaborator named in connection with this specific claim. The “80%+ accuracy” figure appears on the Marici homepage without further substantiation.

What is being predicted? Against which ground-truth dataset is the 80% measured? What is the false-positive rate? Which federal agency, if any, supplies the underlying data, and under what data-sharing instrument? Has the methodology been independently reviewed?

The output of a system marketed under this name is, presumably, intended to be used in or to support a prosecution. In US courts, expert evidence is filtered by the Daubert standard (or its state analogues). Among the factors a court considers are: whether the methodology has been tested, whether it has been subjected to peer review and publication, whether its known or potential error rate is established, and whether it is generally accepted in the relevant scientific community. A novel algorithmic system whose outputs are to be relied on in court would, in the ordinary course, have an associated validation study, a published error-rate disclosure, and a qualified expert capable of testifying to the methodology under cross-examination.

Behavioural inference on identified human subjects — whether perpetrators, witnesses, or survivors — ordinarily falls under research-ethics governance. In the US, that means IRB review (45 CFR 46), informed consent from subjects where feasible, and credentialed supervision of any clinical inference. Where survivor data is involved, additional protections may apply under HIPAA-adjacent frameworks if any health information is processed. A nonprofit running behavioural-scientist inference at the scale Marici implies would ordinarily disclose the IRB of record and the principal investigator.

Operational intelligence products consumed by law-enforcement are ordinarily delivered to a named agency under a memorandum of understanding or a procurement contract. Thorn names its agency users in its annual report and public materials; Polaris Project names the federal contracts under which it operates the US National Human Trafficking Hotline; named agency relationships are the norm rather than the exception in this sector. The agency partner is named because the partnership is the principal substantiation of the work.

For all three named tools: no validation study, no error-rate disclosure, no qualified-expert testimony reference, no court filing that cites the system, no IRB of record, no named principal investigator, no MOU or procurement contract with a named law-enforcement or government partner, and no agency-issued statement confirming use of the system, could be located as of the date of this update.

For “AI Prosecutor”: in how many proceedings, if any, has the system's output been entered or relied upon? Under what error-rate disclosure? For “AI Behavioral Scientist”: which IRB has reviewed the methodology? Which principal investigator is responsible? For “AI Intelligence Analyst”: which agencies are consuming the outputs, and under which MOUs or contracts? Are any of those agencies willing to be named?

This page is a comparison of stated claims against publicly available artefacts. Where evidence may exist privately — under non-disclosure, in unindexed internal documentation, in unreleased agency materials — it has not been disclosed to this publication. Marici has been asked, in the open letter, to make any such evidence available; the full list of questions is reproduced there. Marici's full response will be published on this site, unedited, when received.

Every quotation above is taken verbatim from Marici's own published materials. Every comparator named (Thorn, Polaris Project, OpenAI, Anthropic, Google DeepMind, Meta AI) is a real, identifiable organisation whose own published artefacts are linkable from the sources page or directly searchable.

The conclusion of this page is not that Marici's AI claims are false. The conclusion is that the artefacts which would ordinarily accompany claims of this magnitude could not be located in the public record as of the date of this update. Whether that absence reflects under-publication, confidentiality, or something else is a question only Marici can answer.