Sandery Insights / Professional Intelligence

Research note 04 · Mitch Sandery · August 2026

Garbage In, Generated Out

AI is making job applications better. That may be part of the problem.

Reading time8 minutesEvidence standardResearch + observation

Generative AI can rewrite CVs, tailor applications, identify skills, draft cover letters and increasingly automate the application itself. The output can look excellent.

The more important question is whether the underlying information is good enough.

LinkedIn is a professional directory. It is not a complete career record.

LinkedIn can capture positions, education, skills, certifications, projects, recommendations and other accomplishments.¹ That makes it an enormously useful professional network and discovery layer.

It does not make every profile a full evidentiary account of a career. Responsibility, decision quality, operating constraints, project complexity, progression within roles, failures, relationships and the transferability of experience are often compressed into titles and short descriptions.

A person can accumulate tens of thousands of hours of professional experience while their visible representation amounts to several titles and a few paragraphs. Yet increasingly sophisticated systems are being built on top of that abstraction.

The profile is useful. The mistake is treating it as the person.

AI can enrich language without enriching evidence.

Documented model limitation

Give a language model three weak LinkedIn bullets and it can produce six strong CV bullets. They may be concise, professional and aligned with the vacancy. Unless more evidence has been supplied, the underlying facts have not become richer.

The major model providers are explicit about this limitation. OpenAI defines hallucinations as plausible but false statements and describes them as a continuing challenge.² Anthropic advises users to ground Claude in source material, permit it to express uncertainty and verify claims with direct quotations where accuracy matters.³

Independent testing of generative search tools reinforces the distinction between fluency and provenance. A Tow Center study found frequent citation and attribution failures across eight systems, including ChatGPT Search and Perplexity. That research concerned news retrieval, not CV writing. Its relevance here is narrower: confident synthesis does not guarantee that the underlying evidence is complete or correctly attributed.⁴

Our interpretation: when career inputs are incomplete, a model can improve their representation more reliably than it can recover the missing record.

The output can become more sophisticated than the evidence supporting it.

Users are already describing the failure mode.

Market observation · Not empirical proof

Public reviews and job-search discussions should not be treated as representative samples. They are still useful for identifying recurring product risks.

Users describe general assistants such as ChatGPT and Claude turning thin inputs into generic job-description language, keyword-heavy bullets or achievements that were never supplied.⁸⁹ A review of a specialist AI career platform similarly criticised unrealistically framed bullets that appeared readily identifiable as generated.¹⁰

These observations sit alongside substantial positive feedback for the same category of tools. The point is not that one model or résumé maker is uniformly poor. It is that different tools can converge on the same upstream constraint: they cannot reliably preserve distinctions that the source material never contained.

Observed patternGeneral assistants + specialist tools

Generic language. Invented metrics. Job descriptions presented as achievements.

These are user-reported failure modes, not measured failure rates. They indicate what should be tested and controlled when AI is used to construct professional materials.

The same weakness exists on the other side of the match.

The candidate is often represented by a CV or profile. The opportunity is represented by a job description.

A job description can describe responsibilities, qualifications and broad competencies. It is not necessarily an exhaustive account of the work.⁷ It may say far less about why the role exists, what has already failed, which trade-offs matter, what the organisation is actually like or which situations will distinguish success.

AI matching can therefore become highly sophisticated at comparing two relatively shallow representations. That may improve textual alignment. It does not necessarily improve understanding.

What the system can compareProfile language
against vacancy language

What the decision requiresCareer evidence
against role reality

Then AI optimises both representations.

Candidates can tailor every CV, align keywords, reorder experience and generate polished prose. Individually, this is rational. Collectively, it changes what those signals mean.

If every candidate can tailor well, tailoring becomes less informative. If every candidate can write an excellent cover letter, prose quality becomes less discriminating. If an agent can submit an application, applying itself becomes weaker evidence of intent.

A 2026 survey commissioned by Kickresume found that professional résumé writers were seeing more AI-generated content and identified generic, boilerplate language as their most common issue. The same respondents generally saw AI as a useful tool rather than a threat.⁵ Guidance from the University of Kansas makes the practical point: AI can help with tailoring, but generic or keyword-heavy content may become less effective when a hiring manager reviews it.⁶

01Thin career data

02AI optimisation

03More plausible applications

04Greater filtering

05Still more optimisation

Our interpretation: each participant can become locally more efficient while the market becomes less informationally efficient.

Better automation cannot fully compensate for poor inputs.

Sandery interpretation

CV generation, job discovery, automated applications, sourcing, parsing, ranking and outreach can all create real productivity gains. But there is a ceiling to what downstream optimisation can accomplish if the underlying information remains thin.

Faster processing of incomplete information does not make it complete. A better matching algorithm does not create evidence that was never captured. A larger model cannot reconstruct with certainty what the market never knew.

The more important question may therefore be upstream: what happens if we improve the underlying information before we optimise the transaction?

The competitive advantage may move from generating outputs to owning better inputs.

The CV should be an output. Not the database.

A richer architecture begins with the person rather than the application. It creates a persistent record of experiences, responsibilities, achievements, decisions, operating environments, relationships, complexity, outcomes, progression and demonstrated capabilities.

The CV becomes one representation generated from that underlying intelligence. It can change for a role without changing the source of truth.

The same principle applies to the opportunity. A job description is an input into understanding the role, not necessarily its complete specification. The role can be decomposed into the problems, decisions, capabilities and situations that matter.

Text-matching architecture

Public profile
Generated CV
Job description
Similarity score

Evidence-led architecture

Career evidence
Professional intelligence
Opportunity evidence
Situational match

The Sandery view

Understanding must begin before the application.

Sandery is not being built as another layer of automation on top of LinkedIn. The starting point is a persistent, structured understanding of the individual that is materially deeper than the public profile or current CV.

The Career Bank is intended to preserve that record over time. CV Architect is being developed as the first integrated application: a role-specific document built from the richer source record. Guided CV work is available today through the founding-client programme.

Over time, the same intelligence could support better representation, matching, assessment and professional decisions. The aim is not more information for its own sake. It is better context and judgment around what the information means.

The implication

Better models cannot indefinitely compensate for impoverished inputs.

As generated professional content becomes cheap, the advantage may belong to whoever understands the person best.

Start with the evidence

Build the record behind your CV.

Explore the Career Bank →Explore CV Architect →

References

Evidence and observations used in this note.

The proposed movement of competitive advantage upstream is Sandery interpretation. Product reviews and community discussions illustrate possible failure modes; they are not representative evidence of product quality or prevalence.