Criminals sending phishing emails and fraud messages no longer give themselves away with broken grammar and obvious misspellings. Federal and international cybersecurity authorities have issued a series of warnings confirming that generative AI tools now produce scam text clean enough to fool trained eyes. Since January 2025, the FBI has logged more than 5,100 complaints and $262 million in losses from just one category of impersonation fraud, a scale that reflects how effectively polished, AI-assisted messaging bypasses the defenses people have relied on for years.
Typo-based training is failing against AI-polished fraud
For more than a decade, security awareness programs taught employees and consumers the same shortcut: look for awkward phrasing, misspelled words, and odd formatting. That advice assumed attackers were writing in a second language or copying text carelessly. Generative AI has collapsed that assumption. A recent advisory from the FBI’s Internet Crime Complaint Center warned that criminals now use AI-generated text to “appear believable” and to overcome the common indicators that once helped recipients identify fraud, a shift that directly undercuts traditional training built around spotting obvious language errors.
The practical consequence is direct. Organizations still anchoring their anti-phishing drills to spelling and grammar checks are training people to watch for signals that no longer appear. A reasonable expectation, drawn from the pattern these federal alerts describe, is that within twelve months those organizations will see higher click-through rates on simulated phishing tests than peers who have shifted training toward behavioral and provenance checks, such as verifying sender identity through a separate channel or inspecting URL structures before clicking.
The UK’s National Cyber Security Centre reached a parallel conclusion in its assessment of AI’s near-term impact on cyber threats. In that report, the agency stated that generative AI and large language models will make it difficult for people to assess whether an email or password-reset request is genuine, a finding that specifically targets the surface-level language cues millions of users still depend on. When both U.S. and U.K. cyber authorities warn that polished text is becoming a default feature of criminal outreach, it signals that typo-based skepticism can no longer serve as the first or only line of defense.
FBI and NCSC evidence: scale and speed of AI-enabled scams
The volume of damage is already measurable. The FBI recorded more than 5,100 complaints and $262 million in losses tied to account-takeover fraud through impersonation of financial institution support channels since January 2025. In these schemes, victims received messages, often via text, email, or phone, that appeared to come from their bank’s fraud department. The language was professional, the branding was accurate, and the requests sounded routine. None of the old red flags applied, which made it easier for criminals to pressure victims into sharing one-time passcodes or approving fraudulent transactions.
Separate FBI alerts describe a second front: realistic phishing sites built to mirror employee self-service portals. These fraudulent sites use multichannel notification tactics designed to suppress legitimate security alerts so that victims never see the real warnings their employers send. The combination of flawless written content and convincing web design creates an end-to-end deception chain that no single grammar check can interrupt, especially when messages arrive through text, email, and even voice calls that all reinforce the same false narrative.
The FBI’s San Francisco Field Office reinforced the pattern in a warning tied to outreach around a major cybersecurity conference, noting that attackers are using AI for sophisticated phishing and social engineering attacks across email, voice, and video channels. That release, issued in 2024, served as an early institutional signal that AI-enabled fraud had moved from theoretical risk to active operational threat. It also highlighted that generative tools are not limited to text: audio and video deepfakes can now support email scams by adding apparently “live” confirmations from impostor colleagues or bank representatives.
Taken together, these alerts show a shift in attacker capability that outpaces the defenses most individuals and mid-size organizations have in place. The old model assumed that volume-based scam campaigns would always contain telltale errors because the attackers lacked the time or skill to proofread. Large language models eliminated that bottleneck. A single operator can now generate thousands of unique, grammatically correct messages in minutes, each tailored to a specific target’s context, and can reuse the same AI tools to translate content across languages without introducing the clumsy phrasing that once signaled danger.
Why traditional red flags no longer work
Relying on spelling and grammar as primary warning signs fails for three interconnected reasons. First, generative AI makes it trivial for non-native speakers or low-skilled writers to produce fluent messages, removing the skill gap that defenders had quietly depended on. Second, many legitimate organizations already send poorly written emails, which blurs the distinction between safe and unsafe content and trains users to tolerate mistakes. Third, as attackers adopt AI, they can continually refine their messages based on which versions succeed, creating an evolutionary loop that optimizes for believability rather than speed.
This does not mean that language cues are useless. Some scams still contain inconsistencies, such as mismatched company names or oddly specific demands. But those tells are now the exception, not the rule. Treating them as optional bonuses rather than core screening tools better matches the reality described in recent law-enforcement and national-security assessments.
Gaps in the data and limits of current visibility
Several important questions remain open. No public dataset yet quantifies the share of current phishing or romance scams that actually use AI-generated text versus human-written copy. The FBI’s complaint records do not tag individual reports for the presence or absence of traditional linguistic red flags, which means researchers cannot yet measure exactly how much detection rates have dropped as generative tools spread. And no court filings or convicted-operator statements have surfaced that detail prompt-engineering practices or how success rates changed after adopting large language models.
Financial institutions and email providers have not released longitudinal detection metrics comparing their filtering accuracy before and after widespread AI adoption by attackers. Without that data, the full scope of the problem is still partly estimated rather than precisely measured. The UK assessment frames its findings as a projection through 2025, acknowledging that the threat is evolving faster than measurement infrastructure can track. In practice, this means defenders must make policy and training decisions based on converging warnings and early case studies rather than comprehensive statistics.
Practical steps for readers and organizations
In the absence of perfect data, the safest response is to assume that any unexpected message could be AI-polished and to shift attention from style to origin and behavior. For individuals, that starts with a few habits: never click on links or call phone numbers provided in unsolicited emails or texts; instead, navigate directly to the organization’s official website or use a known contact number. Treat any request for one-time passcodes, remote access, or urgent funds transfer as suspicious, even if the message looks professional and references real account details.
Organizations can adjust training to match this environment. Simulated phishing exercises should include well-written, brand-consistent messages that mirror real business workflows, not just clumsy spoofs. Awareness materials should emphasize verification through independent channels, such as confirming payment changes with a phone call using a number already on file, and should normalize the idea that employees will not be punished for double-checking unusual requests. Technical controls, including multifactor authentication and domain-based message authentication, remain essential but should be framed as backstops rather than guarantees.
Policy makers and regulators, for their part, can encourage more transparent reporting of AI-enabled fraud patterns from major platforms and financial institutions. Standardized fields in complaint databases that capture whether victims noticed language anomalies or interacted with convincingly polished messages would help researchers quantify the problem and refine guidance. Until that visibility improves, the most reliable defense is a mindset shift: assume that criminals can write as clearly as any professional, and focus instead on whether a message’s origin and requested actions truly make sense.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
