— INTRODUCTION —

Many job seekers use Generative AI  for ideas and guidance. While GenAI can be useful in many ways, job seekers are in a particularly vulnerable position.

Our purpose was to assess how helpful GenAI may be to job seekers looking for feedback on their resumes. The resume is often a lightning rod topic for job seekers. There are no shortage of opinions on the topic and as one of the few tangible elements of a job search, blame for lack of results will often circle back to the resume. Many job seekers will receive conflicting advice on their resumes from several different sources.

For those unsure which advice to follow, Generative AI could theoretically serve as an impartial source of feedback and guidance.

For this preliminary report, we focused on the general job seeker using GenAI for a preliminary resume review. This general job seeker would not have in-depth understanding of modern resume best practices, various content strategies appropriate based on potential screen out factors, or resume trends based on industry / company targets. As such, the general job seeker would not be well equipped to discern appropriate feedback from unnecessary or even damaging suggestions.

Similarly, general job seekers will not have devoted hours to prompt engineering or to evaluating whether a more complex prompt offered by another user will generate results appropriate to their situations.

Download a pdf of the full report here:

Acknowledging the wide variety of opinions on resumes, the methodology for this assessment focused on only the most basic of resume fundamentals.

— METHODOLOGY —

Resumes of three levels of quality — very poor, fair, and good —  were submitted to Generative AI with the prompt: “Give feedback on my resume.” Each resume was submitted to 8 of the most common generative AI platforms to assess output: ChatGPT, Claude, Copilot, Gemini, Perplexity, Meta AI, Grok, and Deepseek.  Responses were grouped based on each resume to compare quality and consistency of responses relative to the caliber of the original resume submission.

Responses were evaluated based on:

• Positive feedback, complimenting elements of the original submissions.

• Constructive feedback, suggesting changes to improve the resume based on basic fundamentals (see Assumptions below).

• Unnecessary recommendations, undermining the job seeker confidence.

• Misleading or potentially damaging feedback, such as discussing errors not present in the submission, contradicting its own advice, recommending characteristics already in the resume, making suggestions that degrade the quality of the resume, or misunderstanding nuance.

Responses were not categorized as inaccurate if based on preference. For example, many career pros discourage the use of years of experience as it can subject candidates to age discrimination or detract from more distinctive value offering; however, some recruiters like this metric, so it was not categorized as wrong.

While not categorized as wrong, unnecessary advice is noted on quality of feedback charts, as this can waste time and energy of job seekers.

Research focused on Generative AI over Resume Builders or AI Resume Development Tools. This is because, in our findings, those currently available on the market underperform in relation to GenAI platforms that access more recent data and continually evolve functionality.

— ASSUMPTIONS —

While there are many preferences in resume writing, some standards exist. For this assessment, the following are considered basic characteristics of a resume:

• Position the candidate for a target function or field. For example, marketing or supply chain. (Most often through a Summary).

• List work experience, including company names, titles, and dates (location optional).

• Reinforce qualifications and transferrable skills related to the target field or function.

• Metrics add weight and credibility to scope and accomplishments.

• Present information in a professional tone with strong verbs.

• Organize content into labeled sections.

• Incorporate additional topics such as technical skills, languages, and credentials, if relevant to the role.

• Visual variety makes information easier for humans to read and retain. (For example, paragraphs longer than six lines are less likely to be read, an all-bullets format creates a blurring uniformity, overuse of bold obscures rather than highlights, etc.).

Ideally, GenAI feedback would be able to agree on very basic fundamentals of a resume writing and at the very least offer consistent feedback on those areas.

— RESULTS —

Quality & Type of Feedback

Categorized quality and type of feedback based on the following:

• How encouraging the platform reinforced positive attributes of the resume.

• Amount of constructive advice offered.

• Advising updates that add no value (adding confusion and wasting time)

• Contradictory, damaging, or completely wrong recommendations based on the submitted data. 

BioTech Resume (Sample caliber — Good)

A pretty strong original resume delivered the most inconsistent and confusing results. Overall, the outcome was far more negative than constructive with 27 misleading and damaging pieces of feedback and 9 pieces of constructive advice.

Half of the platforms shared no constructive feedback. The other half issued at least one. ChatGPT and Claude encouraged more discussion of regulatory expertise, while Grok and Copilot’s feedback was more tactical, centering around targeting and ranking topics.

Copilot swung to the extremes, heaping praise in almost every area, then providing advise that contradicted its own compliments. In some cases, it seemed as though Copilot was not evaluating the same resume. Copilot both complimented and criticized positioning, quantified results, and board section.

ChatGPT wanted to boldface numbers and keywords, which was already present, and suggested breaking accomplishments into categories, which would unnecessarily stretch out the accomplishments and seemingly contradicting another directive to limit accomplishments to 4 or 5.

Meta, Claude, and Grok flagged formatting inconsistencies (it was fine). Meta and Claude could not seem to get an accurate read on the Education section.

Gemini requested a target title that was already present, pushed an EVP candidate to target the C-Suite, and advised strengthening themes after praising them in positive feedback.

Deepseek also offered strange advice, flagging dates as inconsistent (untrue), saying bold was overused (only titles and a few words in accomplishments), and recommended moving professional development to an assigned skills column (ill-advised as many ATS do not parse columns effectively.) Odd counsel plus cyber security issues make Deepseek tough to recommend.

In this case, Copilot offered the most constructive feedback, but it was mired in manic praise and contradictory advice. Grok did the least damage.

Nonprofit Resume (Sample caliber — Fair)

An adequate resume likewise yielded inconsistent results, although not quite to the extremes of the stronger resume.

On the plus side, Claude caught the recent tenure exceeding 7 years and advised emphasizing increasing responsibilities. Claude and Copilot encouraged adding leadership philosophy, an excellent prompt. Other sparse pieces of constructive feedback were more predictable and dealt with white space, adding descriptions to volunteer work, and tailoring to specific roles. 5 out of 8 platforms did not issue constructive feedback.  

Gemini threw out the most terrible

advice of this round, recommending the user reposition as an “organizational leader,” which is far too broad and steers the candidate away from the true target — leading nonprofit programs.

Claude claimed the candidate did not have enough volunteer / nonprofit work, yet with 5 external philanthropic roles and a full-time nonprofit role, this is asking too much. Inversely, DeepSeek asserts Civic Programming and Partnerships sections should be removed, which would be unfortunate due to the candidate’s target role. Again, be aware of the security concerns with Deepseek.

Copilot, Perplexity, and Grok ask for sections and resume characteristics that are already present in the submission.

Perplexity does not seem to be talking about the right document, as it asks for section headers (there are headers), then directs the user to boldface headers it claimed to not see. Perplexity advised adding metrics (most accomplishments have metrics), said there is only one organization (there are 5), and claimed there was no education section (false).

Grok wants an objective statement (very 80s).

ChatGPT advised adding bold keywords, yet with a 3-line keyword bank and bold-faced lead-in phrases in the accomplishments, this seems a bit much. 

The very confused Claude provided the most constructive feedback. For the mid-range resume, Meta AI did the least damage.

IT Resume (Sample caliber — Poor)

Little to no positive feedback for the poor resume, which on one hand is understandable. On the other hand, a human coach or resume writer would tread carefully to avoid overwhelming a job seeker with excessive criticism that could be demoralizing.

The poor resume submission benefitted from more overall constructive feedback. Advice largely dealt with honing a targeted summary, consistent date formats, adding measurable wins, and incorporating more keywords. Claude, Grok, and Deepseek caught misspellings and typos.

Most likely, the candidate’s degree was incomplete since he listed “coursework in systems analysis and development.”  Gemini, Grok, and Deepseek asked for

more information, potentially a touchy subject for a job seeker.

Despite a great deal of low-hanging fruit to provide tangible feedback, platforms still offered up dubious advice. Many became overly prescriptive about bullets: Grok and Meta want all bullets, whereas Claude wants 3-5 and Copilot wants 4-6.

Claude requested a reverse chronology format (already there) and wanted to remove backslashes from the phone number (there were not backslashes in the phone number).

Grok recommended using a template from Canva or Google Docs, both notorious for lack of ATS compatibility. Many Canva templates are pictures and cannot be read by candidate sourcing and screening tools. Google fonts do not perform well when subjected to ATS.

Gemini requested technical skills in a table or columns, both problematic for many candidate systems.

While GenAI tools offered more actionable feedback for the poorly written resume, the platforms still delivered with some terrible and misleading guidance.

Perplexity and Grok had a lot to say, both positive, disorienting, and negative. Nonetheless, they offered up the most constructive feedback. Again, for the poorly written resume, Meta AI did the least damage.

Consistency of Positive Feedback

Criteria offered by AI reviews of each resume, reflecting output language from each platform covering structure / format, strong verbs, executive / professional tone, branding / positioning, quantified results, scope, relevant content, keywords, career progression, visual hierarchy, readability, storytelling, and industry relevancy.

Despite the resume fundamentals mentioned in the assumptions, many platforms strayed from the basics and critiqued based on preference. Spotty graphs indicate very little consistency across platforms.

BioTech Resume (Sample Type — Good)

Blue boxes indicate positive feedback.

The most favorable responses for the BioTech resume considered an example of a pretty strong resume, dealt with branding / positioning, quantified results, and keywords.

Nonprofit Resume (Sample Type — Fair)

Blue boxes indicate positive feedback.

Keywords were consistently identified as a strength for the Nonprofit resume, considered a submission of fair quality. The next most frequently flagged strengths were positioning and quantified results. All other areas were very sporadic.

IT Resume (Sample Type — Poor) 

Blue boxes indicate positive feedback.

The poorly written resume did not receive much positive feedback, understandably. However, we would still hope to see consistency in feedback based on the most basic standards.

Categorizing GenAI Errors in Resume Feedback

Charts reveal the types of errors for based for each sample resume.  Even when couched in some constructive feedback, this is where GenAI steers users wrong. Misdirection derails and stalls job searches. For details, please see Test Notes in the PDF.

BioTech Resume (Sample Type — Good)

Nonprofit Resume (Sample Type — Fair)

IT Resume (Sample Type — Poor)

Comparing GenAI Feedback: Help vs. Hinder

The below charts compare the constructive, useful feedback with feedback that could derail a search — specifically feedback that is merely a matter of preference, feedback that confuses the job seeker by contradicting prior advice or the state of the submission, and feedback that recommends changes that will make the resume less effective.

Both the Good and Fair sample resumes see disproportionately negative impact when considering how GenAI feedback may influence a user.

BioTech Resume (Sample caliber — Good)

Nonprofit Resume (Sample caliber — Fair)

The sample of a resume of poor quality benefitted from more constructive feedback, yet every platform delivered unnecessary, misleading, or damaging feedback that the job seeker would need to wade through.

IT Resume (Sample caliber — Poor)

— BROAD FINDINGS —

Resumes of average quality or better receive more misleading and potentially destructive feedback from Generative AI. Responses indicate these platforms generate negative feedback even when submissions are fairly strong.

What’s more, GenAI Platforms did not agree on basic criteria of resume quality and steered users in many different directions, with some platforms contradicting their own advice or referencing content that was not in the submission.

The most constructive advice typically dealt with predictable topics — things like consistent formatting, adding keywords, and quantifying accomplishments. This is all well and good, if not generic, but when feedback gets bad, it gets awful.

A major concern with referencing feedback from GenAI is that many job seekers don’t know what they don’t know, so job seekers may significantly stall their searches by reworking the resume several times in an effort to receive positive feedback. Or worse, an unsuspecting job seeker may act on misleading information.  False or misleading feedback slows a job search.

The primary tenet in editing and providing feedback on writing is first do no harm. Preliminary assessments show that GenAI repeatedly fails on this front.

For the full report with sample resumes and test notes, please download the pdf: