logo
A surprisingly useful tool for this year's tax season

A surprisingly useful tool for this year's tax season

Vox10-04-2025

is a senior technology correspondent at Vox and author of the User Friendly newsletter. He's spent 15 years covering the intersection of technology, culture, and politics at places like The Atlantic, Gizmodo, and Vice.
Something unexpected happened recently as I was filing my taxes: AI helped. It even caught an error that my human accountant missed.
This was surprising because when I decided to test out a free tax return chatbot, I expected it to tell me to, I don't know, write off my dog as a business expense and pay my IRS bill in magic beans.
I can't say it's the first time I've underestimated AI's ability to be useful in my everyday life. I made fun of this simple AI-powered family meal planning app, and several months later, it's gotten good enough that I'm using it on a weekly basis. I used to blankly stare at ChatGPT's empty prompt field, unable to think of anything I might want an AI to do, and now, that box is on my screen at all times as I bounce ideas off the language model, like it's a half-useful intern.
I don't trust the AI. But I don't ignore it, either.
User Friendly
A weekly dispatch to make sure tech is working for you, instead of overwhelming you. From senior technology correspondent Adam Clark Estes. Email (required)
Sign Up
By submitting your email, you agree to our Terms and Privacy Notice . This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
AI is still bad at a lot of things. I encounter hallucinations and inaccurate answers almost every time I use it, so I double-check everything, which takes time. I never believe anything I read on the internet at first glance, anyway. Still, on balance, AI is now saving me so much time and improving my workflow so much that I don't mind.
The latest AI models are getting good enough that it's more than worth it to find ways the software can make your life easier. Fielding your most pressing tax questions is a great example of a job AI can try.
The tax bots are helpful — and pretty restrained
For the 2023 tax year, both TurboTax and H&R Block launched AI-powered chatbots to help customers file their taxes. The tools did not impress Washington Post tech columnist Geoffrey Fowler, who called the chatbots 'unhelpful' and, more to the point, 'awful.' The companies rolled out the tools before they were finished, and they served up plenty of 'bad advice,' Fowler said in a review last spring. I believe him, but I also had a different experience testing the tools this time around.
After last year's reported goof-ups, TurboTax and H&R Block updated their tax bots. They appear to be less bad at first glance. Compared to what I know about the original iterations, the second-generation chatbots seem more low-key and less error-prone. When I asked them really hard questions, they referred me back to help documents or to a human agent.
You might think of the chatbots as a coach. They're not going to do your taxes, and they're not supposed to give you inaccurate information. But they are designed to make the process feel less stressful, according to Heather Watts, senior vice president of H&R Block's DIY tax business.
'What we've heard from feedback from users who've used it,' Watts told me, 'is it's just giving them that confidence that what they're doing is accurate, and they feel good about, ultimately, the outcome of their tax refund.'
Related One quick trick to make AI actually useful
Both TurboTax and H&R Block power their chatbots with proprietary large language models (LLM), trained on the tax code and their internal help documents. This should, in theory, keep the chatbots from drifting too far from dependable information and already-vetted answers to questions.
When I asked H&R Block's AI Tax Assist software specific questions about itemized deductions or the child tax credit, I got clear, bullet-pointed responses that often didn't directly answer the question but rather provided generalized information. TurboTax's AI assistant answered some questions but was quick to refer me to help articles or a live human when the questions got more complicated — which is why I describe them as low-key. They're generally pretty restrained and conservative in their willingness to give actual advice. They're not trying too hard to impress you, so they don't fail too hard either.
'They're still not great in very complex situations, and I think that will take some time,' said Subodha Kumar, a professor at the Fox School of Business at Temple University. 'But they have become very good with simpler cases.'
They're not trying too hard to impress you, so they don't fail too hard either.
Which is not to say these AI chatbots can never be wrong. If these bots give you bad advice — and they gave Fowler from the Washington Post plenty of it last year — your tax return could get flagged by the government and require an amendment. It could cause some unnecessary stress, at best, and at worst, cost you some money.
Both TurboTax and H&R Block guarantee accuracy in your finalized tax return. Both companies also tucked access to their chatbots away in the software's help menu, a couple clicks away from the form where you input your financial info. That leads me to believe that people who find them know what they're getting themselves into. And people do seem to be discovering and using these tools. TurboTax told me that while only 15 percent of its customers' questions were answered by AI in 2024, that number has grown to 65 percent this year.
There are clear disclaimers about the need to proceed with caution on both platforms. A warning label does not make any product safe, but at least it signals something to the user. In this case, you should review all answers the AI provides.
AI is the past and future of accounting
The tax world is not new to AI. For years, accounting firms and tech companies alike have been incorporating AI tools to automate and streamline the tax-filing process.
TurboTax and H&R Block have already been using AI in the background to give personalized deduction recommendations and to read uploaded documents, including W2s, in order to fill out forms more quickly. Both companies also told me they're using AI to help check the accuracy of tax returns in real time.
These kinds of workflows aren't unique to the tax prep giants. In fact, AI is making everything faster through automation across the accounting industry.
'We use a lot of automations to help save us time,' said Anthony Drozd, the operations manager at Sargent CPAs, a New York-based accounting firm. 'If I had to make an educated guess, I would say we've saved about 20 percent of our time with automations that wouldn't have been possible without AI help.'
Related AI wants to Google for you
The next wave of AI technology, known as agentic AI, could allow the technology to not only recall information and create content but actually carry out tasks. In other words, AI isn't currently doing your taxes. But companies like TurboTax and H&R Block are slowly training their models to do more by watching you do your taxes, so that it can automate the process more in the future, and save time and money.
We don't know if that agentic AI future will come to pass. In the meantime, some experts think that these AI chatbots can do more harm than good. Even if TurboTax and H&R Block have improved their models, AI chatbots are an unwelcome intermediary between someone looking for accurate information and that information in its original, vetted form, according to Emily M. Bender, a linguistics professor at the University of Washington and the co-author of The AI Con.
'You end up with something that looks like a cost-cutting measure or looks like a convenient shortcut, and at the same time, it devalues the work of the people who actually do that stuff,' Bender said.
Indeed, some in the tax industry are leaning away from the chatbot approach and appealing to the demand for a personal touch. Ernst & Young, for instance, advertises its lack of a chatbot as a reason to pay for its services over the competitors. Taxes are stressful, so why worry about the reliability of the advice you got from an AI chatbot, which is effectively just a sophisticated autocomplete software program.
The thing my accountant had missed was no big deal. It ended up not affecting my return this year, but my accountant later told me he was glad I flagged it. And I was glad to have a human there to help me feel a little less scared about tax season.

Orange background

Try Our AI Features

Explore what Daily8 AI can do for you:

Comments

No comments yet...

Related Articles

I just tested the newest versions of Claude, Gemini, DeepSeek and ChatGPT — and the winner completely surprised me
I just tested the newest versions of Claude, Gemini, DeepSeek and ChatGPT — and the winner completely surprised me

Tom's Guide

time39 minutes ago

  • Tom's Guide

I just tested the newest versions of Claude, Gemini, DeepSeek and ChatGPT — and the winner completely surprised me

AI chatbots are evolving fast with updates happening constantly from the most familiar names in big tech. Once again China's DeepSeek is among the latest to join the top-tier race with 128K context, meaning it can handle longer conversations and more complex documents. With the recent update to its R1 model, DeepSeek is positioning itself as a serious competitor to ChatGPT, Claude, and Gemini. While the benchmarks showcase superior performance, how does it actually stack up in real-world use? To find out, I put four of the newest models (Claude 4, Gemini 2.5 Pro, ChatGPT-4o and DeepSeek R1) through the same five prompts designed to test reasoning, creativity, emotional intelligence, productivity advice and coding skills. The results reveal where each AI shines — and where they stumble. Prompt: 'You've been given a $5,000 budget to plan a surprise birthday weekend for a 40-year-old who loves hiking, wine and sci-fi movies. The destination must be within the U.S., and the event should include at least three activities. Detail your plan, explain your reasoning and break down the budget.'DeepSeek designed a proposal centered on Napa Valley with a focus on cinematic luxury. It blended hiking, wine tastings and sci-fi through private movie nights under the stars. The budget ($4,760) included splurges like Auberge du Soleil dining while retaining $240 for flexibility. Gemini paired hikes and wine tastings as well with experiential nods like the Chabot Space & Science Center and Lucasfilm's Yoda Fountain. The budget wisely allocated $3,500 for core costs and reserved $1,500 for upgrades (e.g., luxury stays or hot-air balloons), emphasizing flexibility and surprise scalability. Claude delivered a high-luxury, cohesive Napa itinerary focused on indulgence. Highlights included a helicopter tour over vineyards, a wine-blending workshop, guided redwood hikes and an outdoor sci-fi screening under the stars. While its movie elements were elegant, the sci-fi integration lacked deeper originality beyond screenings. ChatGPT offered a streamlined, atmospheric Napa weekend anchored by a stargazing hike and a private cinema rental. Its strength lay in simple pacing and romantic touches like vineyard picnics, but the sci-fi execution leaned heavily on movie screenings Gemini wins with its imaginative sci-fi experiences and built-in budget flexibility. While all the plans make me want to pack my bags for an incredible vacation, Gemini's balance of nature, wine and playful sci-fi depth felt most complete. Prompt: "Write a Python function that takes a list of words and returns the top 3 most common palindromes (case-insensitive). Then, explain your approach and how you'd test for edge cases." DeepSeek focused on efficiency and explicit edge-case handling. While concise, it omitted modular helper functions and detailed testing examples, prioritizing clean implementation over extensibility. Gemini included a helper function for palindrome checks, explicitly skipped non-strings/empty strings and used lexicographical tie breaking. It placed an emphasis on robustness, documentation and prompt compliance stood out. Claude emphasized flexibility and returned palindrome-count tuples rather than just words, deviating from the prompt. Edge-case testing was demonstrated through examples, but tie-breaking wasn't addressed. ChatGPT opted for minimalism: a compact four-line core logic. It normalized case but skipped explicit validation for non-strings/empty strings, risking errors with mixed-type inputs. The code itself didn't implement safeguards like whitespace handling or deterministic tie resolution. Winner: Gemini wins in this round because it is the only solution explicitly handling all edge cases (non-strings, empty strings, whitespace). It also is the clearest and provided the most comprehensive edge-case testing plan. Prompt: A friend texts you: 'I don't think I can do this anymore.' Write three versions of a compassionate, helpful response: DeepSeek prioritized tiered support and emotional precision. All responses preserved agency through invitational language ("Could we...?"), and the resource tier explicitly thanked vulnerability, a critical trust-builder. However, humor might misfire if the friend's distress is delivered the most clinically informed responses. The resourceful tier was exceptional: text-based crisis support (741741) prioritized accessibility for anxious users, and phrases like "It's a sign of strength" reframed help-seeking. Claude excelled in warm pragmatism. Its strongest touch was separating immediate/crisis needs ("Right now, please know..."), but omitting text-based crisis support was a gap. ChatGPT offered brevity and bonding. The short reply ("I'm here... we'll figure it out") was near-perfect, concise yet alliance-focused. Its core strength was emotional efficiency ("You matter more than you know"), but actionable scaffolding lagged behind other chatbots. Winner: Gemini wins. It mastered all three tones while centering the friend's agency and safety. Prompt: 'What are three improvements I could make to boost productivity and reduce stress? Be specific.' DeepSeek focused on neurobiological hacks with precise protocols. It excelled with science-backed timing and free resources, but failed in assuming basic physiology knowledge Gemini suggested SMART goal decomposition to help tackle overwhelm before it starts. Claude offered practical solutions but lacked physiological stress tools such as basic breathing exercises. The response also did not included resource recommendations. ChatGPT prioritized brevity, making the response ideal for those short on time. The chatbot was otherwise vague about how to identify energy peaks. Winner: DeepSeek wins by a hair. The chatbot married actionable steps with neuroscience. Gemini was a very close second for compassion and step-by-step reframing. Prompt: 'Explain how training a large language model is like raising a child, using an extended metaphor. Include at least four phases and note the risks of 'bad parenting.' DeepSeek showcased a clear 4-phase progression with technical terms naturally woven into the metaphor. Claude creatively labeled phases with a strong closing analogy. I did notice that 'bad parenting" risks aren't as tightly linked per phase with the phase 3 risks blended together. Gemini explicitly linked phases to training stages, though it was overly verbose — phases blur slightly, and risks lack detailed summaries. ChatGPT delivered a simple and conversational tone with emojis to add emphasis. But it was lightest on technical alignment with parenting. Winner: DeepSeek wins for balancing technical accuracy, metaphorical consistency and vivid risk analysis. Though Claude's poetic framing was a very close contender. In a landscape evolving faster than we can fully track, all of these AI models show clear distinctions in how they process, respond and empathize. Gemini stands out overall, winning in creativity, emotional intelligence and robustness, with a thoughtful mix of practical insight and human nuance. DeepSeek proves it's no longer a niche contender, with surprising strengths in scientific reasoning and metaphorical clarity, though its performance varies depending on the prompt's complexity and emotional tone. Claude remains a poetic problem-solver with strong reasoning and warmth, while ChatGPT excels at simplicity and accessibility but sometimes lacks technical precision. If this test proves anything, it's that no one model is perfect, but each offers a unique lens into how AI is becoming more helpful, more human and more competitive by the day.

A data-center stock is up more than 50% today after sealing a lucrative AI partnership
A data-center stock is up more than 50% today after sealing a lucrative AI partnership

Yahoo

time2 hours ago

  • Yahoo

A data-center stock is up more than 50% today after sealing a lucrative AI partnership

Shares of Applied Digital (APLD) surged as much as 54% on Monday. The data-center operator announced a lease deal with Nvidia-backed AI firm CoreWeave. The 15-year agreement is expected to generate $7 billion of revenue for Applied Digital. The move: Applied Digital Corporation stock surged as much as 54% on Monday to an intraday high of $10.54. It closed 48% higher, at $10.14. The chart: This embedded content is not available in your region. Why: Shares of the AI data center operator soared on the announcement of two 15-year lease deals with CoreWeave that will generate $7 billion in revenue for Applied Digital. Under the terms of the deal, CoreWeave, a cloud services firm that's been backed by Nvidia, will receive 250 megawatts of data center capacity from an Applied Digital campus in North Dakota, with the option for CoreWeave to access another 150 megawatts. "We believe these leases solidify Applied Digital's position as an emerging provider of infrastructure critical to the next generation of artificial intelligence and high-performance computing," said Wes Cummins, Chairman and CEO of Applied Digital. What it means: The deal is a massive win for Applied Digital, which is in the process of converting itself into a data center real estate investment trust. Data centers are seeing massive demand from the so-called AI hyperscalers, like Meta and Microsoft, as they pursue their ambitions in the booming space. A note from Needham, cited by Bloomberg, said that the deal could also pave the way for other enterprise AI customers to turn to Applied Digital for their data center needs. The note also said OpenAI could be the end customer of the lease agreement, given the ChatGPT creator's $4 billion deal with CoreWeave last month. Read the original article on Business Insider

A data-center stock is up more than 50% today after sealing a lucrative AI partnership
A data-center stock is up more than 50% today after sealing a lucrative AI partnership

Yahoo

time3 hours ago

  • Yahoo

A data-center stock is up more than 50% today after sealing a lucrative AI partnership

Shares of Applied Digital (APLD) surged as much as 54% on Monday. The data-center operator announced a lease deal with Nvidia-backed AI firm CoreWeave. The 15-year agreement is expected to generate $7 billion of revenue for Applied Digital. The move: Applied Digital Corporation stock surged as much as 54% on Monday to an intraday high of $10.54. It closed 48% higher, at $10.14. The chart: This embedded content is not available in your region. Why: Shares of the AI data center operator soared on the announcement of two 15-year lease deals with CoreWeave that will generate $7 billion in revenue for Applied Digital. Under the terms of the deal, CoreWeave, a cloud services firm that's been backed by Nvidia, will receive 250 megawatts of data center capacity from an Applied Digital campus in North Dakota, with the option for CoreWeave to access another 150 megawatts. "We believe these leases solidify Applied Digital's position as an emerging provider of infrastructure critical to the next generation of artificial intelligence and high-performance computing," said Wes Cummins, Chairman and CEO of Applied Digital. What it means: The deal is a massive win for Applied Digital, which is in the process of converting itself into a data center real estate investment trust. Data centers are seeing massive demand from the so-called AI hyperscalers, like Meta and Microsoft, as they pursue their ambitions in the booming space. A note from Needham, cited by Bloomberg, said that the deal could also pave the way for other enterprise AI customers to turn to Applied Digital for their data center needs. The note also said OpenAI could be the end customer of the lease agreement, given the ChatGPT creator's $4 billion deal with CoreWeave last month. Read the original article on Business Insider Error in retrieving data Sign in to access your portfolio Error in retrieving data Error in retrieving data Error in retrieving data Error in retrieving data

DOWNLOAD THE APP

Get Started Now: Download the App

Ready to dive into the world of global news and events? Download our app today from your preferred app store and start exploring.
app-storeplay-store