Comparing the Top AI Chat Services for Resume Reviews
Will artificial intelligence (AI) steal your job? We won’t know that for some time, but for today let’s see how AI can help you get a job. In this article, we’ll compare the top free AI Chat Services and see which delivers the best bang for your (free) buck when asked to improve your resume.
Context
Unless you’ve been living under a rock, you’ve probably heard of ChatGPT — the most popular AI Chat Service known to humanity (as of August 2023, anyways). You may even have read articles about how to use it to write a resume. Going one step further, Resume Builder even has an article entitled: 3 in 4 Job Seekers Who Used ChatGPT to Write Their Resume Got an Interview. Clearly, these services are a game-changer for job seekers.
In this article, we want to take those musings even further and compare the top free AI Chat Services and see which delivers the best result with the least effort.
We’ll compare our AI resume helpers using the following 3 parameters:
- 2 resumes (an ideal resume and one full of errors)
- 3 free AI Chat Services — ChatGPT, Bard, and Claude
- 3 outputs (minimum) per service to check the variability of responses
Based on the results of the above tests, we will crown a winner for best AI resume helper in this article.
About the Authors
Victor received his master’s degree in data science from the University of San Francisco. He has taught courses on data science, machine learning, and low-code data analytics tools such as KNIME. He works as the Director of Data Science Partnerships at the University of San Francisco. His guilty pleasures are deception detection and healthcare data science projects, but his love of the field originally blossomed from natural language processing. Victor is currently based in the United States, where all three AI Chat Services are available.
Mahantesh is a data scientist at KNIME. He has completed his Masters in Data Science from OVGU, Germany. He is actively working in the data science domain in the field of machine learning, natural language processing, deep learning, predictive modeling, and business analytics. He is currently based in Berlin, Germany, where only two of the three AI Chat Services are available — ChatGPT and Bard.
The Resumes
For this experiment, we used two data science resumes — one strong and one weak. The strong resume has the following characteristics:
- No typos.
- Has been used to successfully land jobs for education meets data science roles.
- Has clear, repetitive structure.
- Is metric-oriented, meaning that most bullet points discuss a concrete number. For example, “Developed the target variable and improved baseline roc auc scores from 60% to 89%.”
In contrast, the weak resume has many typos, an extra blank page, differing structure between roles, and does not include a LinkedIn profile. Finally, the weak resume includes irrelevant sections, such as hobbies that do not relate to data science, among many other errors. You can see each resume used below.
The AI Chat Services
Here is what Claude has to say about each AI Chat Service:
ChatGPT is an AI chatbot created by OpenAI that uses a large language model (LLM) to generate human-like conversations on many topics. It was released in November 2022 and quickly gained popularity for its advanced capabilities.
Claude is an AI assistant from Anthropic that uses a limited LLM focused on being helpful, harmless, and honest in its responses. It aims to improve safety in conversational AI.
BARD is Google’s new conversational AI chatbot, announced in February 2023, that utilizes LLMs to compete with ChatGPT. It aims to provide more up-to-date, high-quality responses than previous chatbots.
The Prompt
For each AI Chat Service, we use the same prompt: “Review my resume. Then, suggest improvements. Here is my resume: [resume text]”
The Prompt Output
Strong Resume (SR) & ChatGPT’s Response
Overall, your resume looks impressive, and you have a strong background in data science with relevant experience and education. However, there are some areas where you can make improvements to enhance your resume further:
Formatting and Structure: Consider organizing your resume in a more visually appealing manner to make it easier to read. Use bullet points to break down your accomplishments and responsibilities for each job. Make sure the formatting is consistent throughout.
Summary/Objective Statement: Add a short summary or objective statement at the beginning of your resume to highlight your key skills and goals as a data scientist. This will help recruiters quickly grasp your expertise and what you’re looking for.
Reverse Chronological Order: List your professional experience and education in reverse chronological order, with the most recent positions and degrees first.
Quantify Achievements: Try to quantify your achievements with specific numbers or percentages wherever possible. For example, mention the number of students you taught, the size of the audience for webinars, or the revenue impact of your projects.
Impact-Oriented Language: Use strong action verbs and impact-oriented language to describe your accomplishments. Show how your work positively affected the organizations you were part of.
Remove Personal Information: For privacy reasons, consider removing your full address from the resume. It’s enough to mention your city and state.
Skills Section: Create a separate section for your technical skills and tools. You can categorize them under “Programming Languages,” “Data Analysis,” “Data Visualization,” etc., to provide a clearer overview of your expertise.
Professional Development: If you have participated in any data science-related workshops, conferences, or online courses, consider adding a section to highlight your commitment to continuous learning.
Publications and Research: Mentioning your published work and research projects in a separate section can add weight to your profile. Include your “Machine Learning-based Meal Detection” publication and any other relevant academic contributions.
References: It’s common practice to omit references from the resume. You can mention that references are available upon request.
Next, we’ll look at how those responses differ when we put in the exact same prompt three more times. Here is the first SR ChatGPT reprompt, the second SR ChatGPT reprompt, and the third SR ChatGPT reprompt.
Why reprompt?
AI Chat Service responses are not stable. That is, there is variability using the exact same prompt and in this article we would like to make sure our assessment is a fair one based on several samplings.
On the positive side, most of the recommendations given were some of the most important. For us, quantifying achievements is the single most useful point to mention to anyone writing their resume, and we were happy to see that point mentioned in every single output given by ChatGPT. We did find it odd to mention quantifying achievements for this resume, given that almost every experience had some quantity mentioned. Notice also that in each answer given by ChatGPT, the number of items it suggested to fix were different (10, 11, 12, and 13, respectively). Also of worth, in the third reprompt it gave us a resume that we could copy and paste into a word processor if we chose, which could be useful. Overall, we like the variety of outputs given and would suggest everyone reprompt to see the colorful answers these AI Chat Services can produce.
On the negative side, we found it odd that in the third reprompt it suggested that we remove “Remove Generic Skills” such as Git, SQL, BeautifulSoup, and Selenium. We’re not sure in what world the above would be considered generic skills, and we believe that data scientists should be familiar with version control, know how to deal with databases, and be able to scrape their own data. When we interviewed data scientists, if we didn’t see such skills, we were worried. We also found it odd that it mentioned reverse chronological ordering in the strong resume, even though that was the strategy used throughout. It was either the case that ChatGPT did not understand that the strong resume was in reverse order or it simply defaulted to giving generic (but good) advice. For the other AI Chat Services, we’ll keep on eye on this theme — giving generally useful advice that is irrelevant for the resume being critiqued.
Strong Resume (SR) & Bard’s Response
I can’t assist you with that, as I’m only a language model and don’t have the capacity to understand and respond.
Oddly, when we used Bard in the United States with our given prompt, a response was not given. We retried to make sure it was not a simple bug but to no avail. However, when we used Bard in Europe, it did give 3 distinct responses. Here is the first SR Bard reprompt, the second SR Bard reprompt, and the third SR Bard reprompt.
On the positive side, Bard suggested checking the resume for grammatical and spelling errors, which ChatGPT did not. We also appreciated that only Bard mentioned looking for relevant keywords and putting them in the resume. Keywords help applicants beat the ATS (applicant tracking system) or simple rule-based counting metrics used by companies to quickly filter candidates.
On the negative side, Bard provided some generic suggestions on each section. Like Claude, it was also a bit on the shorter side of descriptions and ideas.
Strong Resume (SR) & Claude’s Response
Here are a few suggestions to improve your resume:
Add a Profile/Summary section at the top that highlights your key skills, years of experience, and specialized expertise areas. This helps provide an overview of your qualifications.
Move Education section to the bottom. Experience and skills are more important.
Expand on your project descriptions by focusing on impact and quantifiable results. For example, how much revenue did you identify for the medical company? How much did your model improvement increase AUC score?
Be more specific about skills used for each project. Don’t just list Python — say Pandas, NumPy, Scikit-Learn, etc.
Remove coursework section under education. Focus on relevant projects instead.
Standardize your formatting and spacing. There are some inconsistencies in dates, indentation, etc.
Consider condensing listing of individual courses. Focus on degrees, major, university name.
Add key achievements under each role — presentations given, automation achieved, skills gained, etc. Quantify reach/impact where possible.
Make skills section more scannable — break into categories like Programming Languages, Data Science Libraries, Databases, etc.
Overall the content looks strong, with a mix of academic and professional projects. With some tweaks to polish up the formatting and organization, it will be a very solid data science resume. Let me know if you would like me to clarify or expand on any of these suggestions.
Next, we’ll look at how those responses differed when we put in the exact same prompt three more times. Here is the first SR Claude reprompt, the second SR Claude reprompt, and the third SR Claude reprompt.
On the positive side, these outputs are shorter in general which can save you time. These shorter responses make the experienced user aware to ask for more or less unlike with ChatGPT which gave us a lot to read with each output. We also enjoyed the specificity of the third reprompt which advised us to quantify and provided us with an example from the resume and then later told us to list out specific tools we used in each role. Finally, the reasoning provided was great for cases like the following: “Remove the coursework section under education. Coursework isn’t as relevant once you have work experience.” Here it was clear that in certain situations you want coursework (but not this one), and we appreciated being taught the reasoning behind the recommendation.
On the negative side, Claude is a bit more terse and tends to give shorter outputs than, say, ChatGPT, so if you didn’t know you can alter the verbosity of the AI Chat Service, you might be left feeling like these responses are lacking. It also tends to give less concrete, actionable material than ChatGPT as well. Comparatively speaking, we would have preferred a bit more verbosity and explanation, which we did find in one of ChatGPT’s responses but did not find as much in Claude’s. Finally, we found it puzzling that three out of four prompts told us to “Move Education section to the bottom”, but the third reprompt told us to do the opposite. In our opinion, put education on the top when you’re a new graduate or the college you went to is prestigious — otherwise, put it on the bottom if you have more relevant work experience. Since a human will only scan your resume for about 6 or 7 seconds, you need to give them the most relevant information first.
Weak Resume (WR) & ChatGPT’s Response
For the weak resume, we’ll just look at three outputs from each service. Here is the first WR ChatGPT output, the second WR ChatGPT output, and the third WR ChatGPT output.
On the positive side, what we enjoyed about the first output was that it gave concrete examples from the weak resume and showed us how to improve it:
For example, instead of: “As a data scientist, We build models for classification, evaluate the results, and share my findings with stakeholders.”
You could write: “Developed machine learning models for classification, resulting in a 20% increase in accuracy compared to previous methods. Presented findings to stakeholders, enabling data-driven decision making and leading to a 15% improvement in business performance.”
This kind of output helps us learn how to improve the rest of the resume and helps us in case we need to update the resume again in the future. It also makes us aware that we need to be looking for projects that have a high impact to put on the resume in the first place! We also liked that in the third output was a fully rewritten (without grammatical or spelling error) version of the weak resume. This critique also felt very tailored to the specific resume it was reading. For instance it commented on the professional experience saying, “Professional Experience (KNIME): Elaborate on the educational materials and blog posts you created. Specify the number of blog posts, the topics covered, and any positive feedback or engagement received from the audience” which is exactly the kind of recommendation we would provide if this were a friend’s resume.
On the negative side, the outputs were not consistent in terms of quality of response and how tailored they were to the resume at hand. In our opinion, the quality of the first output is far better than either the second or third output. At times, it even gave strange feedback, such as the dreaded “References available upon request” line, which should not exist in a modern resume. It also suggested that we “Tailor for Each Job Application,” which isn’t feasible for the domain (data science) given that the number of applications needed to even get an interview is typically in the 200–300s according to the data from our students. Imagine how long it would take to tailor each application? The cost-performance tradeoff simply doesn’t exist. Instead, it should suggest creating 2–3 resumes tailored for a specific sector or data role. Also, note that none of the outputs told me about the spelling or grammatical mistakes we purposely made in the resume.
Weak Resume (WR) & Bard’s Response
For the weak resume, we’ll just look at three outputs from each service. Here is the first WR Bard output, the second WR Bard output, and the third WR Bard output.
On the positive side, it supplied specific recommendations for each section of the weak resume. Find below a sample of the particular suggestions offered by Bard.
- In your professional experience section, you could quantify your accomplishments whenever possible. For example, instead of saying “I built models for classification,” you could say “I built models for classification that achieved an accuracy of 95%.”
- In your education section, you could include the names of any courses that are relevant to the jobs you are applying for.
- In your hobbies and interests section, you could include any activities that demonstrate your skills or interests that are relevant to your professional goals. For example, if you are applying for a job in the healthcare industry, you could mention that you are a volunteer at a local hospital
Also, it did suggest removing the Hobbies section of the resume as it doesn’t add relevance in the context of job application. Another positive aspect of Bard is that it provided a similar response when reprompted, which points towards the stability of the response.
On the negative side, Bard did not point out what spellings should be corrected, it simply suggested getting the resume checked for spelling mistakes. Also, it did not recommend adding LinkedIn or GitHub profile links which can be helpful for technical recruiters or technical round interviews.
Weak Resume (WR) & Claude’s Response
For the weak resume, we’ll just look at three outputs from each service. Here is the first WR Claude output, the second WR Claude output, and the third WR Claude output.
On the positive side, Claude told me to quantify the experiences, which should always be the first recommendation as most resumes I read forget this important detail. And just as with the strong resume, the brevity of responses was appreciated. We also appreciate that Claude can actually read PDFs instead of simply making up content which has been our experience with Bard when given documents.
On the negative side, the output seemed quite generic and didn’t seem as useful compared to output from the other AI Chat Services. As well, for Bard and ChatGPT, key points are bolded so that you can quickly get the gist, whereas with Claude that is not the case. In this sense, the product is not as polished in the presentation of output.
Crowning a Winner
To compare our results, we use a simple infographic with five key terms and stars (maximum of 3 stars and minimum of 1 star). The five key terms are defined below.
Out-of-Box Personalization refers to whether or not the AI Chat Service made specific references to the resumes or simply output generic responses. If we gave more specific prompts, perhaps all services could give more personalized outputs, but Prompt Engineering is beyond the scope of this article.
Stability refers to whether the 4 different outputs were relatively similar or had high variability.
Quality refers to the usefulness of the responses — were the recommendations actually useful or not.
Precision refers to whether the service noted the many mistakes we purposefully made in the weak resume, such as “end-to.end” or whether any service could find a grammatical/spelling flaw in the strong resume.
PDF Function refers to whether the service’s PDF (or other document) reader worked and how well it worked. Note that ChatGPT does not have such functionality. For Bard, the functionality presumably exists but Bard (as of August 2023) usually gives completely unrelated output.
Without further ado, the winner is: ChatGPT! If you’re not interested in prompt engineering and just want a tool to give you a vast amount of ideas to fix your resume with a simple prompt, then ChatGPT is for you.
Final Thoughts
One thing we were pleasantly surprised about was that none of the services suggested we change our names to another nationality which is currently a notorious topic with AI-image tools. Given that a recent attack renders even chat services vulnerable to all sorts of nefarious actors, we were relieved to see that we did not have any such negative outputs. We were also fairly surprised that overall, ChatGPT seemed to give the best responses even though Stanford researchers have seen ChatGPT’s outputs declining in quality recently.
One other surprise to us was that none of the services pointed out specific errors like “end-to.end” (although some did mention checking for spelling errors). The take home message there being, if you are not going to copy paste the resume generated by the AI Chat Service you choose, then you need to ask the service to find and point out all grammatical and spelling errors in the resume explicitly.
On a personal note, Victor has had issues with very complex instructions given to ChatGPT and Bard but noticed that Claude is far better at following complex commands. For instance, if his instructions were long (say 10 distinct commands), ChatGPt and Bard simply ignore some of what is asked. In contrast, Claude can follow very long and complicated instructions well. To us, this work/article suggests that more open-ended, vague, simple questions are better for ChatGPT, but with better prompting, Claude may prove the superior service. We suggest that all users should try out at least 2 services and run their own experiments to see which is better for their particular task.
In the future, it might be helpful to compare the effects of using better prompts to be more specific about exactly what we want from each AI Chat Service. For instance, specifically stating that we want the AI to take the role of a data science manager and seeing if that has any effect on the output. We would also suggest explicitly stating that all grammatical and spelling errors be found in the resume, producing a copy-pasteable version of the resume, and citing sentences from the resume to show how they could be improved. One final instruction of interest might be to ask the AI Chat Service to explain its reasoning so that the resume writer can learn the “why” for all its recommendations. We leave that up to the reader to explore further.
Thanks for reading! May the luck of the LLMs ever be with you.
