Aiberry

Among the most model dependent records in this index. The interview is conducted by a model, the responses are read by models, and the output is entirely model generated. Strip the machine learning out and nothing remains, not even a questionnaire, because the product deliberately refuses the questionnaire format that would otherwise be the fallback.

The architecture is described as an ensemble aggregating three modalities from a single bot administered conversation: what is said, how it is said, and what the face does while saying it. That the assistant itself is animated and conducts the interview means the model occupies both ends of the interaction rather than scoring an artefact a human collected.

This sits above the multimodal assessment vendors that wrap comparable models in documentation, scheduling and engagement layers. There is no workflow suite here to carry value independently, which raises centrality and concentrates risk in the same movement. Every question a buyer has about this vendor is a question about the models, because there is nothing else to ask about.

The clinical product keeps a clinician in the loop by design. The anonymous product structurally cannot, and that is the sharpest unresolved question in this record.

Where a clinician exists, the model is positioned correctly. Smart MindCare produces a depression risk score, symptom level insight and a full transcript into a portal, and publishing the transcript alongside the score is a meaningful oversight feature: a clinician can read what the person actually said rather than accepting a number, which is exactly the reviewability that scoring models usually withhold.

Digital MindCare is the problem. It is anonymous and login free, sold into schools, employers and community health organisations, and the platform is described as returning a risk level assessment for mental health disorders including suicidal ideation. Those two facts together describe a person receiving an indication of suicide risk through a tool that by construction holds no identity, no account and no clinician relationship. Nothing published states what happens next: whether a result is surfaced with crisis resources, whether the organisation is notified and how it could act without an identifier, whether any threshold triggers anything at all, or what a school or employer is instructed to have in place before deployment.

Anonymity is a genuine privacy virtue and it is in direct tension with duty of care here. The tension is not necessarily resolved wrongly, but it is not addressed at all, and a buyer deploying this across a student population is the party who will discover the answer.

No escalation protocol, alerting threshold or human review commitment was located for either product.

Ask what a high risk result shows the person in the anonymous product, what the deploying organisation receives, and what the vendor requires a customer to have in place before go live.

The method is documented in a journal and almost none of it reaches the buyer facing material, which is an unusual shape for this axis.

What is disclosed is genuine. Three modalities are named and the approach is described as an ensemble of machine learners aggregating responses across them. The output is specified more precisely than most: an overall depression risk score, symptom level insight broken out across mood, concentration and energy, and a transcript of every response. Publishing the transcript is itself a transparency feature, because it lets a clinician inspect the input rather than only receiving the verdict. The peer reviewed paper documents the methodology, so a technical reader has somewhere real to go.

The gap sits between the paper and the site. Marketing states clinical equivalence to the gold standard without publishing a single figure. No accuracy, correlation, discrimination or error figure appears on any buyer facing page. A prospect evaluating this product is asked to accept an equivalence claim on trust, when the number exists and is published elsewhere.

The company's own citation compounds it. The research page still links a preprint rather than the 2024 journal article, so a reader following the vendor's link reaches an unreviewed version of work that has since cleared review. The stronger artefact exists and the site points past it.

No model card, architecture description, operating point or threshold is published, and no statement addresses whether third party models underpin the transcription or language components.

Ask for the operating threshold and the sensitivity and specificity at it, and for the research page to cite the published paper.

Not one component of the chain is identified, and the product makes clear that a chain exists.

A dedicated pass located no cloud or infrastructure provider, no base or foundation model, no sub processor register, no third party component list and no position on whether customer data contributes to model development.

The inference available from the product is the reason this reads as an omission rather than a simple absence. A transcript of every response is an advertised feature, so automatic speech recognition is unavoidably in the chain, and it is a component very rarely built in house. The animated conversational assistant that conducts the interview implies a speech or dialogue layer of some kind. Something processes video. Each of those is a place where audio and video of a person discussing their mental health may pass to a party the customer has not been told about, and the company names none of them.

The intellectual lineage is described where the technical stack is not. Public material traces the approach to roughly two decades of brain science research associated with the founding scientist, which explains where the ideas came from and says nothing about what runs in production.

The training question compounds it. Whether interviews recorded in a school, a workplace or a correctional facility feed model development is unaddressed, and the settings involved include populations whose consent is least straightforward.

Ask for the sub processor register, which provider performs transcription, and whether patient audio or video leaves the vendor's own environment at any point.

One real peer reviewed study, correctly disclosed in the journal and misleadingly framed outside it, against almost no commercial evidence.

The study is genuine and it is not a preprint. It appeared in the Journal of Affective Disorders in 2024, a respected psychiatry title, led from the University of Texas at Austin with the Georgetown University Medical Center and the University of Arizona. Nearly 400 participants aged 18 to 74 completed both a Botberry interview and a gold standard depression questionnaire, and the model performed comparably. A published subgroup analysis reporting no evidence of bias by gender, age or race accompanies it. Most vendors in this lane have nothing of this kind, and this record is graded above them for that reason.

The framing around it is where a reader gets misled. Press coverage and the company's own release present a rigorous University of Texas study demonstrating clinical validation. The paper's declarations state it was funded by Aiberry, that two authors are Aiberry employees, and that a third is an Aiberry research coordinator. The employed author carries conceptualisation, methodology, formal analysis and supervision, which is the methodological spine of the work rather than a supporting contribution. The journal disclosure is exactly right; the independent university validation story built on top of it is not what the paper says about itself.

Commercial evidence is close to absent. No customer count, no facility count, no deployment scale figure and no named reference outcome was located in any setting, despite four verticals being marketed. Named collaborators include Advocate Aurora Health, Georgetown and the University of Arizona, with nothing describing what runs where. The company news feed stops in 2023.

Scope is also one condition. Depression is validated. Marketing describes detection of mental health disorders including suicidal ideation, and no study supports that wider claim.

Ask for a deployment outcome at a named site, and for validation covering anything beyond depression.

One genuine architectural minimisation decision, and no description of how anything is actually handled.

The minimisation is real and worth naming plainly, because it is a design choice rather than a policy sentence. Digital MindCare is anonymous and login free by construction. In schools, workplaces and community organisations, which is where volume and reluctance are both highest, the product collects no account and no identified health record at all. A vendor that removes the identifier from the highest volume path has reduced the consequence of a breach rather than promising to guard against one, and that is a stronger form of stewardship than most records here demonstrate.

Handling detail for the identified path is absent. No encryption statement, no key management, no retention schedule, no access control model, no audit logging description, no deletion process and no data ownership position was located for Smart MindCare, which is the product where personal and health information is collected by the company's own description.

The training question is unaddressed in both directions. Nothing states whether recorded interviews contribute to model development, under what consent, or whether an organisation can decline. The material is video and audio of a person describing their mental state, which cannot be de identified the way structured data can, because the face and the voice are the identifier and are simultaneously the input the models read.

The settings sharpen this. Recordings of incarcerated people and of students discussing their mental health are among the most sensitive artefacts any vendor in this index creates.

Ask what is retained from an interview after scoring, for how long, who can retrieve it, and whether recordings train models.

Structurally thoughtful privacy design, resting on a self asserted compliance claim with nothing external behind it.

The structure deserves credit because it is unusual and it is not decorative. Rather than one policy covering everything, three separate policies are published and split by product: one for the anonymous login free screening used by schools, employers and community organisations, one for the account based clinical product where personal and health information is collected, and one for the website. A California resident notice sits alongside them. Splitting the policy by data collection model, rather than writing one document that covers every case vaguely, tells a reader which regime applies to them before they start.

What sits under it is a logo. HIPAA compliance is represented by a badge in the site footer, with no attestation, no assessor, no scope statement and no external verification of any kind located. HIPAA has no certifying body, so a badge is a claim about the company's own view of itself.

Business associate agreement handling is not addressed anywhere. No template, no execution requirement, no statement of which entity contracts and no subcontractor flow down position was located.

One regime specific to the named settings is unaddressed. Behavioral health and recovery centres handle substance use disorder treatment records under a separate federal confidentiality regime with its own consent and redisclosure requirements, and a recovery centre is one of the four settings marketed. Nothing published engages it.

Ask for the business associate agreement template, and how substance use records are handled in the recovery setting.

A single footer logo is the entire published security posture, and it represents the one claim in this area that no external party can attest.

A dedicated pass across the site, its navigation and independent sources located no security page, no trust center, no external attestation of any kind, no penetration testing statement, no vulnerability disclosure policy, no encryption description and no compliance documentation available under agreement. The footer carries a HIPAA badge and nothing accompanies it. HIPAA has no certifying body, so the only security signal published is one that certifies nothing.

The contrast with the same lane is instructive rather than incidental. Vendors handling comparable material publish an attestation report through a public trust center, name their assessor cadence and describe a secure development lifecycle. Buyers in this category increasingly treat an external attestation as an entry requirement rather than a differentiator, and its absence will surface in the first security review at any health system, correctional agency or university.

The data at issue makes the gap more serious than the grade alone conveys. This platform records video and audio of people describing their mental state, including incarcerated people and students. That material is unusually identifiable, unusually sensitive and unusually damaging if exposed, and there is no published evidence that anyone outside the company has examined the controls protecting it.

One pre emptive note: a policy page or a restated compliance claim cannot move this grade. Only an external attestation, or documentation made available under agreement, will.

Ask whether any external security audit has been completed, and what can be shared under a non disclosure agreement.

The position is undeclared rather than overstated, which is a better failure than the alternative and still leaves a buyer without an answer.

Worth noting first what is absent in a good way. This vendor makes no regulatory claim it has not earned. There is no registration badge presented as a quality mark, no designation described as a clearance and no approval language. Against a lane where three separate records carry exactly those overstatements, saying nothing is the more honest silence.

The silence is still total. No device determination, no classification rationale, no clearance, and no statement that the product is positioned outside device regulation was located anywhere.

The classification question is live rather than academic. The output is a depression risk score, and the platform is described as returning a risk level assessment for mental health disorders including suicidal ideation. Decision support treatment generally turns on a clinician being able to review the basis of the recommendation independently, and this product has an unusually strong claim there because it publishes the transcript alongside the score. That helps in the clinical product and does not obviously transfer to the anonymous product, where no clinician sits between the model and the person receiving a risk result.

Two settings carry their own regimes that nothing addresses. Screening in correctional facilities and screening students in higher education each raise consent and mandated reporting questions distinct from device law.

Ask for the written device determination, and whether it covers both products or only the clinician mediated one.

Published subgroup analysis in a peer reviewed journal, which is more than nearly any record in this index offers, carrying a statistical caveat the company does not state.

The substance is real. The 2024 study reports no evidence of bias by gender, age or race, and the company describes its training and validation population as reflecting the composition of the United States population, spanning ages 18 to 74 and depression severity from minimal to severe. The positioning is specific rather than general: the company contrasts itself with platforms relying on vocal biomarkers alone, which it identifies as commonly showing gender bias, and multimodality is offered as the mitigation. That is a testable claim about a named competing approach and it is unusual to see one made.

The caveat is what the finding can carry. A null result in roughly 400 participants is an underpowered null, not a demonstration of fairness. Once that sample is divided across racial categories, sexes and age bands, the study can detect only large disparities; moderate ones would be invisible. Absence of evidence of bias and evidence of absence of bias are different findings, and the marketing language treats them as the same one.

Three specific gaps remain. No subgroup discrimination figures are published, only the null. Skin tone is not analysed, which is the variable that drives differential error in facial analysis and is not interchangeable with self reported race. And primary language is not examined at all, though the models read word choice and speech patterns.

A recruitment claim also appears that the study does not support: pipelines described as designed to eliminate bias including by sexual orientation, a category the published analysis does not cover.

Graded B because a published subgroup analysis is a real disclosure and outranks the principles pages graded lower here. Ask for subgroup performance figures rather than the null, and for a skin tone analysis.

Responsibility is unallocated, and the product architecture makes the question harder here than for a conventional clinical tool.

A dedicated pass located no service level agreement, no accuracy warranty, no performance guarantee, no indemnity, no remediation commitment and no published statement of recourse. Terms of use sit behind the application login rather than on the public site, so the contractual position cannot be examined before contact.

The published evidence does not fill the gap and should not be read as filling it. Clinical equivalence to a questionnaire is a research finding about average behaviour across a sample, not a commitment about an individual result, and because no operating threshold is published there is no stated false negative rate for the risk score.

What distinguishes this record is who absorbs a miss. In the clinical product a clinician stands between the score and the patient and carries professional responsibility. In the anonymous product there is no clinician, no identity and no account, so a missed indication of suicide risk reaches a person with nobody positioned to notice, and the deploying school or employer has no way to know it happened. That is not an ordinary vendor liability question, and no public material acknowledges it as one.

One pre emptive note: additional published research, a further validation study or a wider customer list cannot move this grade. Only a contractual term, or a published operating point carrying a stated false negative rate, will.

Ask what the customer agreement says about responsibility for a missed risk indication, and whether the anonymous product is contracted on different terms.

Not one record system is named, and the single integration claim made names nothing either.

A dedicated pass located no electronic health record integration, no interface standard, no marketplace or partner listing, no application programming interface documentation and no data exchange description. The only integration statement found asserts that the corporate wellness version integrates easily with all leading corporate wellness platforms, which names none of them and covers the one setting where a clinical record system is least likely to be involved.

The absence weighs more heavily here than it would for a wellness product, because two of the four marketed settings are clinical. Behavioral health and recovery centres run on record systems, and a screening instrument whose results cannot reach the chart produces an outcome measure that lives in a separate portal and has to be transcribed by hand to become part of the clinical record. The company's own product description reinforces the concern: results are presented as landing in an Aiberry portal, described as a central hub holding all patient scores, which is a destination rather than a connection.

A correctional or university deployment raises the same question against different systems, and neither is addressed.

Graded D consistent with how a total absence of named systems is treated across this index. Ask whether any record system integration exists in production today, through what standard, and whether a risk score can be written to a chart as a discrete result.

The delivery model can be inferred and is never stated.

What can be established comes from observation rather than disclosure: a browser based application served from a company subdomain, with a login for organisational users and a separate anonymous path requiring no account. That is a multi tenant cloud product by every visible sign, and the company nowhere says so.

Everything a buyer would need is absent. No cloud provider, no hosting region, no data residency commitment, no tenancy model, no segregation description, no backup or disaster recovery position and no customer controlled deployment option was located.

Two of the marketed settings make this more consequential than a bare omission usually is. Correctional agencies are government buyers who ordinarily impose their own hosting, authorisation and data location requirements, and public universities frequently carry state level data residency obligations. Nothing published indicates whether either can be accommodated, and a procurement team in those settings will hit this question in the first security review rather than the last.

The anonymous product raises a distinct question in the same area. Screening without an account still generates video and audio that must be transmitted and processed somewhere, and where that happens, and what is retained after a score is produced, is unstated.

Ask where the platform is hosted, in which region, under what tenancy model, and whether a government or public institution deployment can be isolated.

Cost is absent from every published surface. A dedicated pass located no pricing page, no unit of charge, no range, no implementation fee position, no minimum commitment and no pilot terms. Navigation offers a demo booking form as the only commercial path, and the get started page is a feature list terminating in the same form.

Two products with structurally different economics make the silence more costly. Digital MindCare is anonymous and login free, which implies volume based or site based charging since there are no seats to count. Smart MindCare is account based with a clinician portal, which implies per clinician or per patient charging. Nothing indicates which, or whether an organisation buying one can move to the other.

Four buyer settings compound it further. A correctional facility, a university counselling service, a corporate wellness administrator and a recovery centre have entirely different budget structures, procurement routes and volume profiles, and a single undisclosed model cannot be assumed to serve all four.

No return proxy is offered either. The commercial argument is clinician time saved against a provider shortage, and no cost per assessment, time released figure or throughput number is published to support it.

Ask for the unit of charge for each product, how the anonymous product is metered without accounts, and the minimum commitment by setting.

Breadth of setting is real and breadth of validated clinical scope is not, and the gap between the two is the finding on this axis.

Four settings are named and each is a genuinely distinct deployment context: corrections, corporate wellness, behavioral health and recovery, and higher education. Reaching into correctional facilities and universities is unusual in this index and addresses populations that conventional behavioral health vendors do not target. The anonymous login free product is well matched to the school and workplace contexts where an identified health record would deter participation.

Validated clinical scope is one condition. The published study covers depression in adults aged 18 to 74. Marketing describes screening for mental health disorders including suicidal ideation, and no validation supports either the wider disorder set or a risk assessment for suicidal ideation specifically. A buyer in a correctional or university setting is being sold a breadth the evidence does not yet cover.

Two populations sit inside the named settings and outside the validated range. Higher education includes students under 18 in some programmes, and the study population starts at 18. Correctional populations carry consent and coercion considerations that nothing published addresses, and a screening instrument administered where participation may not be freely refused is a materially different product from the same instrument in a clinic.

Nothing addresses languages other than English, which matters for a model reading speech patterns and word choice, and nothing addresses markets outside the United States.

Ask for the validated age range in practice, what supports the suicidal ideation claim, and the consent model in corrections.